From c6cb50ecb7eb09be78340fe4d6f8dd43c53a372c Mon Sep 17 00:00:00 2001
From: lin-bot23 <linzhixiong23@gmail.com>
Date: Mon, 26 Oct 2026 18:06:23 +0800
Subject: [PATCH 01/13] docs(minimax): document H3 camera-motion vocabulary and
 turbo losses on fast action

---
 .../video/minimax/minimax-h3-prompt-guide.mdx | 27 ++++++++++++++++---
 es/tutorials/video/minimax/minimax-h3.mdx     | 13 ++++-----
 .../video/minimax/minimax-h3-prompt-guide.mdx | 27 ++++++++++++++++---
 ja/tutorials/video/minimax/minimax-h3.mdx     |  7 ++---
 .../video/minimax/minimax-h3-prompt-guide.mdx | 27 ++++++++++++++++---
 ko/tutorials/video/minimax/minimax-h3.mdx     |  7 ++---
 .../video/minimax/minimax-h3-prompt-guide.mdx | 23 +++++++++++++++-
 tutorials/video/minimax/minimax-h3.mdx        |  3 ++-
 .../video/minimax/minimax-h3-prompt-guide.mdx | 27 ++++++++++++++++---
 zh/tutorials/video/minimax/minimax-h3.mdx     |  7 ++---
 10 files changed, 139 insertions(+), 29 deletions(-)

diff --git a/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
index a40b226850..74de258a6d 100644
--- a/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
+++ b/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
@@ -2,13 +2,13 @@
 title: "Guía de prompts de ComfyUI MiniMax H3: guías oficiales, consejos y embeddings de prompt"
 description: "Escribe mejores prompts de MiniMax H3 en ComfyUI: guías oficiales, consejos generales, reglas de diálogo y hablantes, y embeddings de estilo."
 sidebarTitle: "Guía de prompts"
-translationSourceHash: 0d858ae8
+translationSourceHash: 0d517d21
 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx
 translationBlockHashes:
   "_intro": da22cc9d
   "Official guides": 3d762c67
   "General tips": 91822299
-  "Shots and cuts": 8184dae6
+  "Shots and cuts": 8fc80fd4
   "Dialogue and speakers": 4985438e
   "Prompt embeddings": 9c899169
 ---
@@ -59,7 +59,28 @@ Numera las tomas `[Shot 1]`, `[Shot 2]`, … `[Shot 1]` empieza en 0.00s. Las to
 - Usa una sola marca de tiempo (`At 00:05.000`), no un rango como `[5s-8s]`
 - Formula los cortes como acciones: `the camera cuts to`, `the shot transitions to`, etc. Usa disolvencias, desvanecidos o barridos cuando haga falta
 
-Inicia una toma nueva cuando cambien el sujeto, el espacio, el estado, el punto de vista o el momento. Los reencuadres pequeños siguen siendo una sola toma con movimiento de cámara.
+Un corte introduce información nueva sobre el sujeto, el espacio, el estado, el punto de vista o el momento. El estado también puede cambiar dentro de una toma continua. Cuando solo cambia la distancia o un ángulo leve, mantén una sola toma y mueve la cámara.
+
+### Movimiento de cámara
+
+Escribe el movimiento como una acción dentro de la toma: primero el tipo de movimiento, después la amplitud y la velocidad.
+
+```text
+[Shot 1] ... The camera pushes in with small amplitude at slow speed toward the letter in her hands.
+[Shot 2] At 00:03.500, the camera pans right with large amplitude at fast speed, revealing the open doorway.
+```
+
+| Elemento | Valores |
+| --- | --- |
+| Tipo de movimiento | `Zoom In` / `Zoom Out`, `Push In` / `Pull Out`, `Pan Left` / `Pan Right`, `Truck Left` / `Truck Right`, `Tilt Up` / `Tilt Down`, `Pedestal Up` / `Pedestal Down`, `Arc Shot`, `Tracking Shot`, `Static Shot`, `Shake Slightly` / `Shake Strongly`, `POV`, `Roll Clockwise` / `Roll Counterclockwise` |
+| Amplitud | `with small amplitude`, `with large amplitude` |
+| Velocidad | `at slow speed`, `at fast speed` |
+
+`Zoom In` / `Zoom Out` cambian la distancia focal y el cuerpo de la cámara no se mueve; `Push In` / `Pull Out` desplazan la cámara.
+
+La amplitud solo admite small y large, y la velocidad solo slow y fast. Omite ambas cuando quieras amplitud media y ritmo normal.
+
+Escribe el movimiento dentro de la acción. No acumules las palabras como etiquetas al final de la frase.
 
 ## Diálogo y hablantes
 
diff --git a/es/tutorials/video/minimax/minimax-h3.mdx b/es/tutorials/video/minimax/minimax-h3.mdx
index cf96b9cb2c..dc79b448f7 100644
--- a/es/tutorials/video/minimax/minimax-h3.mdx
+++ b/es/tutorials/video/minimax/minimax-h3.mdx
@@ -2,7 +2,7 @@
 title: "Guía de generación de video con MiniMax H3 en ComfyUI"
 description: "Usa MiniMax H3 de pesos abiertos en ComfyUI: flujos de texto a video, imagen a video y referencia a video con audio estéreo nativo y Sage Attention."
 sidebarTitle: "Descripción general"
-translationSourceHash: 6baf68cb
+translationSourceHash: eb426452
 translationFrom: tutorials/video/minimax/minimax-h3.mdx
 translationBlockHashes:
   "_intro": 94f54ee8
@@ -10,12 +10,12 @@ translationBlockHashes:
   "Getting started": eed90954
   "Choosing the checkpoint precision": 4a53fdf9
   "Workflow index": bdc3b9f8
-  "Setting the output resolution": db770c15
-  "Clip length": d9104665
-  "Step count": d2a8a6f8
+  "Setting the output resolution": 1fda252d
+  "Clip length": 2d495d82
+  "Step count": 184c83d7
   "Sampler and scheduler": 79c739c3
   "Speeding up generation with Sage Attention": 24cc9a70
-  "Speeding up generation with sparse attention": cd054e87
+  "Speeding up generation with sparse attention": 801c9acf
   "Commercial licensing": 201e2de5
 ---
 import UpdateReminder from "/snippets/es/tutorials/update-reminder.mdx";
@@ -156,7 +156,7 @@ Valores predeterminados (Lightning LoRA **desactivado**): **20 pasos**. Con turb
 
 Las tomas de referencia necesitan más pasos: los tokens de referencia viajan en cada paso, así que el turbo de 4 pasos apenas puede aplicar las referencias (deriva de pose y rostro). Para un ajuste preciso de la referencia, mantén el turbo desactivado o aumenta hacia **25**.
 
-Los schedules destilados suavizan el detalle fino (piel, tela, cabello) frente a la base de 20 pasos. Cada archivo turbo es específico de un modo; prueba con una semilla fija antes de apilar LoRAs de estilo. `turbo_steps` y `turbo_model_strength` son ajustables: más pasos + menor fuerza (menos de 1) te acercan de nuevo a la calidad base y consumen tiempo. Ejemplo: archivo de 4 pasos a 6–8 pasos, fuerza ~0.8.
+Los schedules destilados suavizan el detalle fino (piel, tela, cabello) frente a la base de 20 pasos. Cada archivo turbo es específico de un modo; prueba con una semilla fija antes de apilar LoRAs de estilo. `turbo_steps` y `turbo_model_strength` son ajustables: más pasos + menor fuerza (menos de 1) te acercan de nuevo a la calidad base y consumen tiempo. Ejemplo: archivo de 4 pasos a 6–8 pasos, fuerza ~0.8. Quita turbo y usa los 20 pasos completos cuando la acción rápida o el movimiento de cámara importen más que la velocidad.
 
 Los LoRA turbo acortan el schedule, no el tiempo por paso (se fusionan al cargar). Ejecutar un LoRA de 4 pasos a 20 pasos tarda aproximadamente lo mismo que la base de 20 pasos.
 
@@ -164,6 +164,7 @@ Orientación según el contenido:
 
 - Tomas simples: **12–16** pasos suelen ser suficientes
 - Detalle de alta frecuencia (cota de malla, filigrana, muchos objetos pequeños): mejora hasta ~**50**; por debajo de eso, aparecen artefactos de cuadrícula triangular en movimiento (peor con turbo)
+- Acción rápida y movimiento de cámara: turbo con 4–8 pasos suaviza ambos, y la acción rápida puede verse más lenta de lo descrito. Un plano estático lo oculta, así que compara recuentos de pasos en un clip con movimiento real
 - La mayor parte de la mejora en movimiento/prompt se logra al paso **16**; hay poco cambio visible después de **50**
 - Los fotogramas suaves en resolución nativa suelen ser un problema de resolución, no de pasos
 
diff --git a/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
index ad079feee8..3e0eb3cf76 100644
--- a/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
+++ b/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
@@ -2,13 +2,13 @@
 title: "ComfyUI MiniMax H3 プロンプトガイド: 公式ガイド、ヒント、プロンプト埋め込み"
 description: "ComfyUI でより良い MiniMax H3 プロンプトを書く方法: MiniMax の公式プロンプト作成ガイド、一般的なプロンプトのコツ、セリフと話者のルール、スタイル埋め込み。"
 sidebarTitle: "プロンプトガイド"
-translationSourceHash: 0d858ae8
+translationSourceHash: 0d517d21
 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx
 translationBlockHashes:
   "_intro": da22cc9d
   "Official guides": 3d762c67
   "General tips": 91822299
-  "Shots and cuts": 8184dae6
+  "Shots and cuts": 8fc80fd4
   "Dialogue and speakers": 4985438e
   "Prompt embeddings": 9c899169
 ---
@@ -60,7 +60,28 @@ non_diegetic_music: ...
 - 単一時刻（`At 00:05.000`）、範囲不可
 - `the camera cuts to` など; 必要ならディゾルブ/フェード
 
-被写体・空間・状態・視点・時間が変わるときに新ショット。微リフレームは 1 ショットでカメラモーション。
+カットは被写体・空間・状態・視点・時間について新しい情報を導入します。状態の変化は 1 つの連続ショット内でも起こせます。距離やわずかな角度だけが変わるときは 1 ショットのままカメラを動かします。
+
+### カメラモーション
+
+カメラの動きはショット内の動作文として書きます。モーションの種類、続けて振幅と速度です。
+
+```text
+[Shot 1] ... The camera pushes in with small amplitude at slow speed toward the letter in her hands.
+[Shot 2] At 00:03.500, the camera pans right with large amplitude at fast speed, revealing the open doorway.
+```
+
+| 要素 | 値 |
+| --- | --- |
+| モーションの種類 | `Zoom In` / `Zoom Out`、`Push In` / `Pull Out`、`Pan Left` / `Pan Right`、`Truck Left` / `Truck Right`、`Tilt Up` / `Tilt Down`、`Pedestal Up` / `Pedestal Down`、`Arc Shot`、`Tracking Shot`、`Static Shot`、`Shake Slightly` / `Shake Strongly`、`POV`、`Roll Clockwise` / `Roll Counterclockwise` |
+| 振幅 | `with small amplitude`、`with large amplitude` |
+| 速度 | `at slow speed`、`at fast speed` |
+
+`Zoom In` / `Zoom Out` は焦点距離が変わり、カメラ本体は動きません。`Push In` / `Pull Out` はカメラ自体が前後に移動します。
+
+振幅は small と large のみ、速度は slow と fast のみです。中程度の振幅と通常の速度でよいときは両方省略します。
+
+動きは動作文に書き込みます。文末にラベルとして並べないでください。
 
 ## セリフと話者
 
diff --git a/ja/tutorials/video/minimax/minimax-h3.mdx b/ja/tutorials/video/minimax/minimax-h3.mdx
index 97bd3e5f18..08246a424c 100644
--- a/ja/tutorials/video/minimax/minimax-h3.mdx
+++ b/ja/tutorials/video/minimax/minimax-h3.mdx
@@ -2,7 +2,7 @@
 title: "ComfyUI MiniMax H3 動画生成ガイド"
 description: "ComfyUI でオープンウェイトの MiniMax H3 を使う方法: テキストから動画、画像から動画、参照から動画のワークフローとネイティブステレオオーディオ、プロンプト作成のコツ、Sage Attention による高速化。"
 sidebarTitle: "概要"
-translationSourceHash: e95e545e
+translationSourceHash: eb426452
 translationFrom: tutorials/video/minimax/minimax-h3.mdx
 translationBlockHashes:
   "_intro": 94f54ee8
@@ -12,7 +12,7 @@ translationBlockHashes:
   "Workflow index": bdc3b9f8
   "Setting the output resolution": 1fda252d
   "Clip length": 2d495d82
-  "Step count": d2a8a6f8
+  "Step count": 184c83d7
   "Sampler and scheduler": 79c739c3
   "Speeding up generation with Sage Attention": 24cc9a70
   "Speeding up generation with sparse attention": 801c9acf
@@ -157,7 +157,7 @@ H3 は短クリップで学習; ネイティブノードは広い範囲を受け
 
 参照ショットはステップ多め: 参照トークンは毎ステップ; 4 ステップ turbo は参照がほぼ乗らず（ポーズ/顔ドリフト）。厳密な参照一致では turbo オフまたは **25** 付近。
 
-蒸留スケジュールは 20 ステップ基線より細部が滑らか。各 turbo ファイルはモード別; スタイル LoRA を重ねる前に固定シードで試験。`turbo_steps` 増 + `turbo_model_strength` 低下（&lt;1）で基線品質に近づき時間増。例: 4 ステップファイルを 6–8 ステップ、強度約 0.8。
+蒸留スケジュールは 20 ステップ基線より細部が滑らか。各 turbo ファイルはモード別; スタイル LoRA を重ねる前に固定シードで試験。`turbo_steps` 増 + `turbo_model_strength` 低下（&lt;1）で基線品質に近づき時間増。例: 4 ステップファイルを 6–8 ステップ、強度約 0.8。速い動作やカメラモーションを速度より優先するなら、turbo を外して 20 ステップで回します。
 
 Turbo は schedule を短くするだけで 1 ステップは速くなりません（読み込み時マージ）。4 ステップ LoRA を 20 ステップ回すと基線 20 と同程度の時間。
 
@@ -165,6 +165,7 @@ Turbo は schedule を短くするだけで 1 ステップは速くなりませ
 
 - 単純ショット: **12–16** ステップで足りることが多い
 - 高周波詳細（チェインメイル、フィリグリー、小物体多数）: 約 **50** まで改善; それ未満は動き中の三角グリッドアーティファクト（turbo で悪化）
+- 速い動作とカメラモーション: turbo の 4–8 ステップではどちらも弱まる。速い動作は指定より遅く見えることもある。静止ショットでは差が出ないため、ステップ数の比較は実際に動くクリップで行う
 - モーション/プロンプトの大半は **16** ステップまで; **50** 以降は差が小さい
 - ネイティブ解像度のソフトさは多くが解像度問題
 
diff --git a/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
index 4970df9ad9..8c9278194f 100644
--- a/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
+++ b/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
@@ -2,13 +2,13 @@
 title: "ComfyUI MiniMax H3 프롬프트 가이드: 공식 가이드, 팁, 프롬프트 임베딩"
 description: "ComfyUI에서 더 나은 MiniMax H3 프롬프트 작성하기: MiniMax 공식 프롬프트 작성 가이드, 일반 프롬프트 팁, 대사와 화자 규칙, 스타일 임베딩."
 sidebarTitle: "프롬프트 가이드"
-translationSourceHash: 0d858ae8
+translationSourceHash: 0d517d21
 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx
 translationBlockHashes:
   "_intro": da22cc9d
   "Official guides": 3d762c67
   "General tips": 91822299
-  "Shots and cuts": 8184dae6
+  "Shots and cuts": 8fc80fd4
   "Dialogue and speakers": 4985438e
   "Prompt embeddings": 9c899169
 ---
@@ -60,7 +60,28 @@ non_diegetic_music: ...
 - 단일 시각(`At 00:05.000`), 구간 아님
 - `the camera cuts to` 등; 필요 시 디졸브/페이드
 
-피사체·공간·상태·시점·시간 변경 시 새 샷. 미세 리프레임은 한 샷·카메라 모션.
+컷은 피사체·공간·상태·시점·시간에 대한 새 정보를 전달합니다. 상태 변화는 하나의 연속 샷 안에서도 일어날 수 있습니다. 거리나 약간의 각도만 바뀔 때는 한 샷으로 두고 카메라를 움직입니다.
+
+### 카메라 모션
+
+카메라 움직임은 샷 안의 동작 문장으로 씁니다. 모션 종류, 이어서 진폭과 속도입니다.
+
+```text
+[Shot 1] ... The camera pushes in with small amplitude at slow speed toward the letter in her hands.
+[Shot 2] At 00:03.500, the camera pans right with large amplitude at fast speed, revealing the open doorway.
+```
+
+| 요소 | 값 |
+| --- | --- |
+| 모션 종류 | `Zoom In` / `Zoom Out`, `Push In` / `Pull Out`, `Pan Left` / `Pan Right`, `Truck Left` / `Truck Right`, `Tilt Up` / `Tilt Down`, `Pedestal Up` / `Pedestal Down`, `Arc Shot`, `Tracking Shot`, `Static Shot`, `Shake Slightly` / `Shake Strongly`, `POV`, `Roll Clockwise` / `Roll Counterclockwise` |
+| 진폭 | `with small amplitude`, `with large amplitude` |
+| 속도 | `at slow speed`, `at fast speed` |
+
+`Zoom In` / `Zoom Out`은 초점 거리가 바뀌고 카메라 본체는 고정입니다. `Push In` / `Pull Out`은 카메라가 직접 이동합니다.
+
+진폭은 small과 large만, 속도는 slow와 fast만 있습니다. 중간 진폭과 보통 속도면 둘 다 생략합니다.
+
+움직임은 동작 문장에 넣어 씁니다. 문장 끝에 라벨로 나열하지 마세요.
 
 ## 대사와 화자
 
diff --git a/ko/tutorials/video/minimax/minimax-h3.mdx b/ko/tutorials/video/minimax/minimax-h3.mdx
index 9b862322a8..46417a3934 100644
--- a/ko/tutorials/video/minimax/minimax-h3.mdx
+++ b/ko/tutorials/video/minimax/minimax-h3.mdx
@@ -2,7 +2,7 @@
 title: "ComfyUI MiniMax H3 비디오 생성 가이드"
 description: "ComfyUI에서 오픈 가중치 MiniMax H3를 사용하는 방법: 텍스트 기반, 이미지 기반, 레퍼런스 기반 비디오 워크플로와 네이티브 스테레오 오디오, 프롬프트 작성 팁, Sage Attention 가속화."
 sidebarTitle: "개요"
-translationSourceHash: e95e545e
+translationSourceHash: eb426452
 translationFrom: tutorials/video/minimax/minimax-h3.mdx
 translationBlockHashes:
   "_intro": 94f54ee8
@@ -12,7 +12,7 @@ translationBlockHashes:
   "Workflow index": bdc3b9f8
   "Setting the output resolution": 1fda252d
   "Clip length": 2d495d82
-  "Step count": d2a8a6f8
+  "Step count": 184c83d7
   "Sampler and scheduler": 79c739c3
   "Speeding up generation with Sage Attention": 24cc9a70
   "Speeding up generation with sparse attention": 801c9acf
@@ -157,7 +157,7 @@ H3는 짧은 클립으로 학습되었습니다. 네이티브 노드는 더 넓
 
 참조 샷은 스텝 더 필요: 참조 토큰 매 스텝; 4스텝 turbo는 참조 거의 미적용(포즈/얼굴 드리프트). 밀착 참조는 turbo 끄거나 **25** 근처.
 
-증류 schedule은 20스텝 기준보다 디테일 평활. turbo 파일은 모드별; 스타일 LoRA 전 고정 시드 테스트. `turbo_steps`↑ + `turbo_model_strength`↓(&lt;1)로 기본 품질·시간 trade-off. 예: 4스텝 파일 6–8스텝, 강도 ~0.8.
+증류 schedule은 20스텝 기준보다 디테일 평활. turbo 파일은 모드별; 스타일 LoRA 전 고정 시드 테스트. `turbo_steps`↑ + `turbo_model_strength`↓(&lt;1)로 기본 품질·시간 trade-off. 예: 4스텝 파일 6–8스텝, 강도 ~0.8. 빠른 동작이나 카메라 모션이 속도보다 중요하면 turbo를 끄고 20스텝으로 돌립니다.
 
 Turbo는 schedule만 단축, 스텝당 시간 동일(로드 시 병합). 4스텝 LoRA 20스텝 ≈ 기본 20스텝 시간.
 
@@ -165,6 +165,7 @@ Turbo는 schedule만 단축, 스텝당 시간 동일(로드 시 병합). 4스텝
 
 - 단순 샷: **12–16** 스텝 종종 충분
 - 고주파 디테일(체인메일, 필리그리, 작은 물체 다수): ~**50**까지; 미만 시 움직임 중 삼각 격자 아티팩트(turbo worse)
+- 빠른 동작과 카메라 모션: turbo 4–8 스텝에서는 둘 다 약해지고, 빠른 동작은 묘사보다 느리게 보일 수 있습니다. 정지 샷에서는 차이가 드러나지 않으므로 스텝 수 비교는 실제로 움직이는 클립에서 합니다
 - 모션/프롬프트 이득 대부분 **16** 전; **50** 이후 차이 작음
 - 네이티브 해상도 soft는 대개 해상도 문제
 
diff --git a/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
index 4e910e92d1..e130811923 100644
--- a/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
+++ b/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
@@ -51,7 +51,28 @@ Number shots `[Shot 1]`, `[Shot 2]`, … `[Shot 1]` starts at 0.00s. Later shots
 - Use a single timestamp (`At 00:05.000`), not a range like `[5s-8s]`
 - Phrase cuts as actions: `the camera cuts to`, `the shot transitions to`, etc. Dissolves/fades/wipes when needed
 
-New shot when subject, space, state, viewpoint, or time changes. Small reframes stay one shot with camera motion.
+A cut introduces new information about the subject, space, state, viewpoint, or time. State can also change inside one continuous shot. When only the distance or a slight angle changes, keep one shot and move the camera.
+
+### Camera motion
+
+Write the move as an action inside the shot: motion type, then amplitude and speed.
+
+```text
+[Shot 1] ... The camera pushes in with small amplitude at slow speed toward the letter in her hands.
+[Shot 2] At 00:03.500, the camera pans right with large amplitude at fast speed, revealing the open doorway.
+```
+
+| Dimension | Values |
+| --- | --- |
+| Motion type | `Zoom In` / `Zoom Out`, `Push In` / `Pull Out`, `Pan Left` / `Pan Right`, `Truck Left` / `Truck Right`, `Tilt Up` / `Tilt Down`, `Pedestal Up` / `Pedestal Down`, `Arc Shot`, `Tracking Shot`, `Static Shot`, `Shake Slightly` / `Shake Strongly`, `POV`, `Roll Clockwise` / `Roll Counterclockwise` |
+| Amplitude | `with small amplitude`, `with large amplitude` |
+| Speed | `at slow speed`, `at fast speed` |
+
+`Zoom In` / `Zoom Out` change the focal length while the camera body stays still; `Push In` / `Pull Out` move the camera itself.
+
+Small and large are the only amplitude values, slow and fast the only speeds. Omit both when the medium range and normal pacing are what you want.
+
+Write the move into the action. Do not stack the words as labels at the end of the sentence.
 
 ## Dialogue and speakers
 
diff --git a/tutorials/video/minimax/minimax-h3.mdx b/tutorials/video/minimax/minimax-h3.mdx
index 8741491142..d15e44f238 100644
--- a/tutorials/video/minimax/minimax-h3.mdx
+++ b/tutorials/video/minimax/minimax-h3.mdx
@@ -142,7 +142,7 @@ Defaults (Lightning LoRA **off**): **20 steps**. With turbo on: **8** for T2V/I2
 
 Reference shots need more steps: reference tokens ride every step, so 4-step turbo can barely apply references (pose/face drift). For tight reference match, keep turbo off or raise toward **25**.
 
-Distilled schedules smooth fine detail (skin, fabric, hair) vs 20-step base. Each turbo file is mode-specific; test at a fixed seed before stacking style LoRAs. `turbo_steps` and `turbo_model_strength` are adjustable: higher steps + lower strength (less than 1) move back toward base quality and cost time. Example: 4-step file at 6–8 steps, strength ~0.8.
+Distilled schedules smooth fine detail (skin, fabric, hair) vs 20-step base. Each turbo file is mode-specific; test at a fixed seed before stacking style LoRAs. `turbo_steps` and `turbo_model_strength` are adjustable: higher steps + lower strength (less than 1) move back toward base quality and cost time. Example: 4-step file at 6–8 steps, strength ~0.8. Drop turbo and run the full 20 steps when fast motion or camera movement matters more than speed.
 
 Turbo LoRAs shorten the schedule, not per-step time (merged at load). Running a 4-step LoRA at 20 steps takes about as long as base 20 steps.
 
@@ -150,6 +150,7 @@ Guidance by content:
 
 - Simple shots: **12–16** steps often enough
 - High-frequency detail (chainmail, filigree, many small objects): improves up to ~**50**; below that, triangular grid artifacts under motion (worse with turbo)
+- Fast action and camera moves: turbo at 4–8 steps softens both, and fast action can read slower than described. Static shots hide it, so compare step counts on a clip with real movement
 - Most motion/prompt gains by step **16**; little visible change past **50**
 - Soft frames at native resolution are usually a resolution issue, not steps
 
diff --git a/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
index 2b69cde3df..09a26b1b22 100644
--- a/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
+++ b/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
@@ -2,13 +2,13 @@
 title: "ComfyUI MiniMax H3 提示词指南：官方指南、技巧与提示词嵌入"
 description: "在 ComfyUI 中写出更好的 MiniMax H3 提示词：MiniMax 官方提示词编写指南、通用提示词技巧、对白与说话人规则，以及风格嵌入。"
 sidebarTitle: "提示词指南"
-translationSourceHash: 0d858ae8
+translationSourceHash: 0d517d21
 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx
 translationBlockHashes:
   "_intro": da22cc9d
   "Official guides": 3d762c67
   "General tips": 91822299
-  "Shots and cuts": 8184dae6
+  "Shots and cuts": 8fc80fd4
   "Dialogue and speakers": 4985438e
   "Prompt embeddings": 9c899169
 ---
@@ -60,7 +60,28 @@ non_diegetic_music: ...
 - 写单点时间（`At 00:05.000`），不写区间
 - 用 `the camera cuts to` 等动作描述；需要时可写溶解/淡入淡出
 
-仅景别/微角度变化时保持同一镜头，用运镜表达。
+切换用来引入主体、空间、状态、视角或时间的新信息，状态也可以在一个连续镜头内变化。只有距离或轻微角度变化时，保持同一镜头，用运镜完成。
+
+### 运镜
+
+运镜写在镜头内的动作句里：先写运动方式，再写幅度和速度。
+
+```text
+[Shot 1] ... The camera pushes in with small amplitude at slow speed toward the letter in her hands.
+[Shot 2] At 00:03.500, the camera pans right with large amplitude at fast speed, revealing the open doorway.
+```
+
+| 维度 | 取值 |
+| --- | --- |
+| 运动方式 | `Zoom In` / `Zoom Out`、`Push In` / `Pull Out`、`Pan Left` / `Pan Right`、`Truck Left` / `Truck Right`、`Tilt Up` / `Tilt Down`、`Pedestal Up` / `Pedestal Down`、`Arc Shot`、`Tracking Shot`、`Static Shot`、`Shake Slightly` / `Shake Strongly`、`POV`、`Roll Clockwise` / `Roll Counterclockwise` |
+| 幅度 | `with small amplitude`、`with large amplitude` |
+| 速度 | `at slow speed`、`at fast speed` |
+
+`Zoom In` / `Zoom Out` 改变焦距、机身不动；`Push In` / `Pull Out` 是机身本身在移动。
+
+幅度只有 small 和 large 两个取值，速度只有 slow 和 fast。要中等幅度、常速时，两者都省略。
+
+把运镜写进动作句。不要把标签堆在句尾。
 
 ## 对白与说话人
 
diff --git a/zh/tutorials/video/minimax/minimax-h3.mdx b/zh/tutorials/video/minimax/minimax-h3.mdx
index 71ee2b5e31..b0bea164ff 100644
--- a/zh/tutorials/video/minimax/minimax-h3.mdx
+++ b/zh/tutorials/video/minimax/minimax-h3.mdx
@@ -2,7 +2,7 @@
 title: "ComfyUI MiniMax H3 视频生成指南"
 description: "如何在 ComfyUI 中使用开放权重的 MiniMax H3：文生视频、图生视频、参考生视频工作流，支持原生立体声音频，附提示词编写技巧与 Sage Attention 加速方法。"
 sidebarTitle: "概述"
-translationSourceHash: e95e545e
+translationSourceHash: eb426452
 translationFrom: tutorials/video/minimax/minimax-h3.mdx
 translationBlockHashes:
   "_intro": 94f54ee8
@@ -12,7 +12,7 @@ translationBlockHashes:
   "Workflow index": bdc3b9f8
   "Setting the output resolution": 1fda252d
   "Clip length": 2d495d82
-  "Step count": d2a8a6f8
+  "Step count": 184c83d7
   "Sampler and scheduler": 79c739c3
   "Speeding up generation with Sage Attention": 24cc9a70
   "Speeding up generation with sparse attention": 801c9acf
@@ -157,7 +157,7 @@ H3 在短片段上训练；节点接受更宽范围且不阻止超长请求。
 
 参考镜头需更多步数：参考 token 每步都参与，4 步 turbo 可能几乎贴不上参考（姿态/面部漂移）。需紧贴参考时关闭 turbo 或提高到约 **25**。
 
-蒸馏步数比 20 步基线更平滑细节。各 turbo 文件按模式训练；叠风格 LoRA 前请固定种子测试。可提高 `turbo_steps` 并降低 `turbo_model_strength`（小于 1）换回基线质感。例：4 步文件用 6–8 步、强度约 0.8。
+蒸馏步数比 20 步基线更平滑细节。各 turbo 文件按模式训练；叠风格 LoRA 前请固定种子测试。可提高 `turbo_steps` 并降低 `turbo_model_strength`（小于 1）换回基线质感。例：4 步文件用 6–8 步、强度约 0.8。动作或运镜比速度更重要时，关掉 turbo，跑满 20 步。
 
 Turbo 缩短的是 schedule，不是单步耗时（加载时合并）。4 步 LoRA 跑 20 步约等于基线 20 步耗时。
 
@@ -165,6 +165,7 @@ Turbo 缩短的是 schedule，不是单步耗时（加载时合并）。4 步 Lo
 
 - 简单镜头：**12–16** 步常够
 - 高频细节（链甲、花纹、大量小物体）：可改善至约 **50**；更低易出现运动下三角网格伪影（turbo 更严重）
+- 快速动作与运镜：turbo 在 4–8 步会削弱这两项，快速动作还可能比描述更慢。静态镜头看不出差异，对比步数要在有真实运动的片段上看
 - 运动/提示词收益多在 **16** 步前；**50** 步后差异很小
 - 原生分辨率下画面发软多为分辨率问题，而非步数
 

From 7d9a94e102de0076f6f0ea9367a431a7cbcd030e Mon Sep 17 00:00:00 2001
From: lin-bot23 <linzhixiong23@gmail.com>
Date: Tue, 27 Oct 2026 18:11:10 +0800
Subject: [PATCH 02/13] docs(minimax): document the audio-first pass for H3

---
 es/tutorials/video/minimax/minimax-h3.mdx | 23 +++++++++++++++++------
 ja/tutorials/video/minimax/minimax-h3.mdx | 17 ++++++++++++++---
 ko/tutorials/video/minimax/minimax-h3.mdx | 17 ++++++++++++++---
 tutorials/video/minimax/minimax-h3.mdx    | 12 +++++++++++-
 zh/tutorials/video/minimax/minimax-h3.mdx | 17 ++++++++++++++---
 5 files changed, 70 insertions(+), 16 deletions(-)

diff --git a/es/tutorials/video/minimax/minimax-h3.mdx b/es/tutorials/video/minimax/minimax-h3.mdx
index dc79b448f7..030994fc66 100644
--- a/es/tutorials/video/minimax/minimax-h3.mdx
+++ b/es/tutorials/video/minimax/minimax-h3.mdx
@@ -2,10 +2,10 @@
 title: "Guía de generación de video con MiniMax H3 en ComfyUI"
 description: "Usa MiniMax H3 de pesos abiertos en ComfyUI: flujos de texto a video, imagen a video y referencia a video con audio estéreo nativo y Sage Attention."
 sidebarTitle: "Descripción general"
-translationSourceHash: eb426452
+translationSourceHash: fa293eed
 translationFrom: tutorials/video/minimax/minimax-h3.mdx
 translationBlockHashes:
-  "_intro": 94f54ee8
+  "_intro": 71cca49e
   "Key features": 8ca92280
   "Getting started": eed90954
   "Choosing the checkpoint precision": 4a53fdf9
@@ -13,6 +13,7 @@ translationBlockHashes:
   "Setting the output resolution": 1fda252d
   "Clip length": 2d495d82
   "Step count": 184c83d7
+  "Generating the soundtrack first": 7f97b544
   "Sampler and scheduler": 79c739c3
   "Speeding up generation with Sage Attention": 24cc9a70
   "Speeding up generation with sparse attention": 801c9acf
@@ -24,7 +25,7 @@ import UpdateReminder from "/snippets/es/tutorials/update-reminder.mdx";
 
 La documentación se reparte en seis páginas:
 
-- **Descripción general** (esta página): capacidades, checkpoints, índice de flujos de trabajo, resolución, duración del clip, pasos, sampler y aceleraciones
+- **Descripción general** (esta página): capacidades, checkpoints, índice de flujos de trabajo, resolución, duración del clip, pasos, audio primero, sampler y aceleraciones
 - **[Flujos de trabajo nativos](/es/tutorials/video/minimax/minimax-h3-native)**: T2V, I2V, R2V y técnicas con nodos nativos
 - **[Referencia multifotograma](/es/tutorials/video/minimax/minimax-h3-multiframe)**: fija fotogramas de referencia en la línea de tiempo
 - **[Fun ControlNet Union](/es/tutorials/video/minimax/minimax-h3-fun-controlnet)**: controla el video o haz relleno inteligente con una máscara
@@ -170,13 +171,23 @@ Orientación según el contenido:
 
 El audio converge más tarde que el video en el denoise conjunto. El habla y el timbre necesitan **12+** pasos; con **8** pasos el audio suele ser el eslabón débil.
 
+## Generar primero la banda sonora
+
+El audio y el video se generan en el mismo denoise, así que los schedules cortos y los LoRAs turbo dañan el audio antes que la imagen. Dos pasadas los separan: fija el audio en un lienzo pequeño y después condiciona el video con él.
+
+1. Haz una pasada de texto a video del clip a **32 × 32**, con turbo desactivado, **40–50** pasos y la misma duración que tendrá el plano final. Los fotogramas latentes de audio dependen de la duración del clip, no del tamaño del lienzo, así que la pasada produce una pista completa por un coste cercano al de un solo paso a resolución completa. Desconecta Resolution Selector e introduce el tamaño en los `width` y `height` que alimentaba; su **Megapixels** solo baja hasta **0.1**. De esta pasada solo se conserva el audio.
+2. Decodifícalo con `VAEDecodeAudio` y guárdalo con `SaveAudioAdvanced`.
+3. En el flujo R2V, añade `LoadAudio`, conéctalo a `ref_audio` y deja `audio_vae` conectado. Cita la pista como `<Audio 1>` e indica la relación: `fully_copy` conserva el audio de referencia y `reference` solo toma la voz, en cuyo caso el diálogo lo escribes tú (ver la [guía de prompts](/es/tutorials/video/minimax/minimax-h3-prompt-guide)). La pasada de video trabaja entonces con una voz y un ritmo ya fijados en lugar de inventar habla en 4 a 8 pasos.
+
+El tamaño del lienzo intercambia calidad de voz por tiempo. **32 × 32** es casi gratis, pero la voz puede adoptar un tono robótico. **256 × 256** acerca el timbre a un render a resolución completa. Por encima, la pasada cuesta tanto como el render al que alimenta, así que sube solo cuando la voz tenga que ser de calidad final. El habla y las letras suelen sobrevivir al lienzo pequeño; los efectos y la música ligados a la acción en pantalla no, así que genéralos en la pasada final.
+
 ## Sampler y scheduler
 
 Las plantillas usan **`res_multistep`** + **`simple`**. T2V/I2V/FastH3 los ocultan dentro del subgrafo; R2V, Multiframe y Fun ControlNet exponen `KSamplerSelect` y `BasicScheduler` en el lienzo.
 
 `res_multistep` es multistep de segundo orden (el primer paso es Euler). Dividir la programación entre dos samplers descarta el historial multistep: el primer paso del segundo sampler vuelve a ser de primer orden. Mantén una sola ejecución del sampler, o dale al segundo suficientes pasos para calentar.
 
-Los desplazamientos H3 predeterminados son `shift` **12** y `audio_shift` **3** según la definición del modelo (sin nodo adicional). Las plantillas FastH3 configuran **MiniMax H3 Sigma Shift** con `shift_video` **10**, `shift_audio` **3**. Un valor de shift de video incorrecto en versiones destiladas (p. ej., **3** en lugar de **10** en 8 pasos) puede mostrar artefactos de cuadrícula. Los finetunes personalizados pueden necesitar su propio par en **MiniMax H3 Sigma Shift** (consulta la tarjeta de modelo del archivo). [LoRA DMAD de 4 pasos](https://huggingface.co/ZhengmingYu/DMAD/blob/main/minimax_h3/dmad_minimax_h3_4step_full_critic_comfyui.safetensors) usa `shift` **12**, `audio_shift` **2**, **`lcm`** + **`simple`**.
+Los desplazamientos H3 predeterminados son `shift` **12** y `audio_shift` **3** según la definición del modelo (sin nodo adicional). Las plantillas FastH3 configuran **ModelSamplingMiniMaxH3** con `shift_video` **10**, `shift_audio` **3**. Un valor de shift de video incorrecto en versiones destiladas (p. ej., **3** en lugar de **10** en 8 pasos) puede mostrar artefactos de cuadrícula. Los finetunes personalizados pueden necesitar su propio par en **ModelSamplingMiniMaxH3** (consulta la tarjeta de modelo del archivo). [LoRA DMAD de 4 pasos](https://huggingface.co/ZhengmingYu/DMAD/blob/main/minimax_h3/dmad_minimax_h3_4step_full_critic_comfyui.safetensors) usa `shift` **12**, `audio_shift` **2**, **`lcm`** + **`simple`**.
 
 ## Acelerar la generación con Sage Attention
 
@@ -196,11 +207,11 @@ La deformación al final del clip o el texto distorsionado en pantalla con Sage
 - Las plantillas usan checkpoints **`int8_convrot`**. **Comfy Kitchen no les da soporte** ([problema #15529](https://github.com/Comfy-Org/ComfyUI/issues/15529)). Usa la atención predeterminada o cambia `UNETLoader` a **`pruned_bf16`** y luego cambia el backend.
 - En checkpoints **bf16**, la **atención de Comfy Kitchen** (rotación de canales antes de cuantizar) elimina los artefactos a una velocidad similar.
 
-Añade **Backend de atención del modelo** (`comfy kitchen attention`) en la ruta del modelo hacia `BasicGuider` (después de LoRA/turbo), o usa `--use-ck-attention` (0.32.0+). Requiere los kernels INT8 de `comfy-kitchen` en tu hardware.
+Añade **Model Attention Backend** (`comfy kitchen attention`) en la ruta del modelo hacia `BasicGuider` (después de LoRA/turbo), o usa `--use-ck-attention` (0.32.0+). Requiere los kernels INT8 de `comfy-kitchen` en tu hardware.
 
 ## Acelerar la generación con atención dispersa
 
-**Atención dispersa por bloques del modelo** (0.35.0+, experimental) reduce el costo de atención en clips largos. Para H3 base, usa `method` **`sol-attn`** (`sla`/`vsa` necesitan pesos correspondientes). Requiere CUDA + `sol_attn` de `comfy-kitchen`; de lo contrario, las capas permanecen densas.
+**Model Sparse Attention** (0.35.0+, experimental) reduce el costo de atención en clips largos. Para H3 base, usa `method` **`sol-attn`** (`sla`/`vsa` necesitan pesos correspondientes). Requiere CUDA + `sol_attn` de `comfy-kitchen`; de lo contrario, las capas permanecen densas.
 
 - Valores predeterminados: `start_percent` **0.2**, `end_percent` **1.0**. Un inicio más tardío (~**0.4**) y un final más temprano (~**0.9**) protegen los fotogramas de apertura y cierre.
 - Mantén **`sink_conditioning`** en **`exact_kv_and_rows`** en H3 (protege las filas empaquetadas de texto, audio y referencia).
diff --git a/ja/tutorials/video/minimax/minimax-h3.mdx b/ja/tutorials/video/minimax/minimax-h3.mdx
index 08246a424c..fb17ae3041 100644
--- a/ja/tutorials/video/minimax/minimax-h3.mdx
+++ b/ja/tutorials/video/minimax/minimax-h3.mdx
@@ -2,10 +2,10 @@
 title: "ComfyUI MiniMax H3 動画生成ガイド"
 description: "ComfyUI でオープンウェイトの MiniMax H3 を使う方法: テキストから動画、画像から動画、参照から動画のワークフローとネイティブステレオオーディオ、プロンプト作成のコツ、Sage Attention による高速化。"
 sidebarTitle: "概要"
-translationSourceHash: eb426452
+translationSourceHash: fa293eed
 translationFrom: tutorials/video/minimax/minimax-h3.mdx
 translationBlockHashes:
-  "_intro": 94f54ee8
+  "_intro": 71cca49e
   "Key features": 8ca92280
   "Getting started": eed90954
   "Choosing the checkpoint precision": 4a53fdf9
@@ -13,6 +13,7 @@ translationBlockHashes:
   "Setting the output resolution": 1fda252d
   "Clip length": 2d495d82
   "Step count": 184c83d7
+  "Generating the soundtrack first": 7f97b544
   "Sampler and scheduler": 79c739c3
   "Speeding up generation with Sage Attention": 24cc9a70
   "Speeding up generation with sparse attention": 801c9acf
@@ -25,7 +26,7 @@ import UpdateReminder from "/snippets/ja/tutorials/update-reminder.mdx";
 
 ドキュメントは 6 ページに分かれています:
 
-- **概要**（このページ）: 機能、チェックポイント、ワークフロー索引、解像度、クリップ長、ステップ、サンプラー、高速化
+- **概要**（このページ）: 機能、チェックポイント、ワークフロー索引、解像度、クリップ長、ステップ、音声先行、サンプラー、高速化
 - **[ネイティブワークフロー](/ja/tutorials/video/minimax/minimax-h3-native)**: T2V、I2V、R2V とネイティブノードのテクニック
 - **[マルチフレーム参照](/ja/tutorials/video/minimax/minimax-h3-multiframe)**: タイムライン上の参照フレームを固定
 - **[Fun ControlNet Union](/ja/tutorials/video/minimax/minimax-h3-fun-controlnet)**: コントロールビデオまたはマスクによるインペイント
@@ -171,6 +172,16 @@ Turbo は schedule を短くするだけで 1 ステップは速くなりませ
 
 音声は共同デノイズで映像より収束が遅い。音声・音色は **12+** ステップ; **8** では音声が弱いことが多い。
 
+## サウンドトラックを先に生成する
+
+映像と音声は同じデノイズで生成されるため、短いスケジュールや turbo では映像より先に音声が崩れます。2 パスに分ければ切り離せます。まず小さなキャンバスで音声を確定し、その音声を条件にして映像を生成します。
+
+1. このクリップをテキストから動画で 1 回生成します。**32 × 32**、turbo オフ、**40–50** ステップ、長さは最終カットと同じにします。音声 latent のフレーム数はクリップ長だけで決まり、キャンバスサイズには依存しないため、このパスでも全長の音声が得られ、コストはフル解像度の 1 ステップ程度です。Resolution Selector のリンクを外し、それが接続していた `width` と `height` にサイズを入力してください。このノードの **Megapixels** は **0.1** が下限です。残すのは音声だけです。
+2. `VAEDecodeAudio` でデコードし、`SaveAudioAdvanced` でファイルに保存します。
+3. R2V ワークフローで `LoadAudio` を追加してそのファイルを読み込み、`ref_audio` に接続します（`audio_vae` は接続したまま）。プロンプトでは `<Audio 1>` として参照し、関係を明記します。`fully_copy` は参照した音声をそのまま使い、`reference` は声質だけを借ります。その場合はセリフを自分で書きます（[プロンプトガイド](/ja/tutorials/video/minimax/minimax-h3-prompt-guide)参照）。これで映像側は、4–8 ステップで音声を作り直すのではなく、確定した声とタイミングに合わせて動きます。
+
+キャンバスサイズは音質と時間のトレードオフです。**32 × 32** はほぼ無コストですが、声に機械的な響きが出ることがあります。**256 × 256** なら音色はフル解像度のレンダーに近づきます。それ以上にすると、このパスのコストは供給先のレンダーと同程度になるため、音声を最終品質にしたいときだけ上げてください。セリフや歌詞は小さなキャンバスでも大抵成立しますが、画面の動きに紐づく効果音やスコアは成立しないため、最終パスで生成します。
+
 ## サンプラーとスケジューラ
 
 テンプレートは **`res_multistep`** + **`simple`**。T2V/I2V/FastH3 はサブグラフ内; R2V、マルチフレーム、Fun ControlNet はキャンバスに `KSamplerSelect` と `BasicScheduler`。
diff --git a/ko/tutorials/video/minimax/minimax-h3.mdx b/ko/tutorials/video/minimax/minimax-h3.mdx
index 46417a3934..0d4da36fee 100644
--- a/ko/tutorials/video/minimax/minimax-h3.mdx
+++ b/ko/tutorials/video/minimax/minimax-h3.mdx
@@ -2,10 +2,10 @@
 title: "ComfyUI MiniMax H3 비디오 생성 가이드"
 description: "ComfyUI에서 오픈 가중치 MiniMax H3를 사용하는 방법: 텍스트 기반, 이미지 기반, 레퍼런스 기반 비디오 워크플로와 네이티브 스테레오 오디오, 프롬프트 작성 팁, Sage Attention 가속화."
 sidebarTitle: "개요"
-translationSourceHash: eb426452
+translationSourceHash: fa293eed
 translationFrom: tutorials/video/minimax/minimax-h3.mdx
 translationBlockHashes:
-  "_intro": 94f54ee8
+  "_intro": 71cca49e
   "Key features": 8ca92280
   "Getting started": eed90954
   "Choosing the checkpoint precision": 4a53fdf9
@@ -13,6 +13,7 @@ translationBlockHashes:
   "Setting the output resolution": 1fda252d
   "Clip length": 2d495d82
   "Step count": 184c83d7
+  "Generating the soundtrack first": 7f97b544
   "Sampler and scheduler": 79c739c3
   "Speeding up generation with Sage Attention": 24cc9a70
   "Speeding up generation with sparse attention": 801c9acf
@@ -25,7 +26,7 @@ import UpdateReminder from "/snippets/ko/tutorials/update-reminder.mdx";
 
 문서는 6개 페이지로 나뉩니다:
 
-- **개요**(이 페이지): 기능, 체크포인트, 워크플로 인덱스, 해상도, 클립 길이, 스텝, 샘플러, 가속
+- **개요**(이 페이지): 기능, 체크포인트, 워크플로 인덱스, 해상도, 클립 길이, 스텝, 오디오 우선, 샘플러, 가속
 - **[네이티브 워크플로](/ko/tutorials/video/minimax/minimax-h3-native)**: T2V, I2V, R2V 및 네이티브 노드 기법
 - **[멀티프레임 레퍼런스](/ko/tutorials/video/minimax/minimax-h3-multiframe)**: 타임라인에 레퍼런스 프레임 고정
 - **[Fun ControlNet Union](/ko/tutorials/video/minimax/minimax-h3-fun-controlnet)**: 컨트롤 비디오 또는 마스크 인페인팅
@@ -171,6 +172,16 @@ Turbo는 schedule만 단축, 스텝당 시간 동일(로드 시 병합). 4스텝
 
 오디오는 공동 디노이즈에서 영상보다 늦게 수렴. 음성·음색 **12+**; **8**에서 오디오 약함.
 
+## 사운드트랙 먼저 생성하기
+
+영상과 오디오는 같은 디노이즈에서 함께 생성되므로, 스텝이 짧거나 turbo를 쓰면 영상보다 오디오가 먼저 무너집니다. 두 번의 패스로 분리할 수 있습니다. 작은 캔버스에서 오디오를 먼저 확정하고, 그 오디오를 조건으로 영상을 생성합니다.
+
+1. 이 클립을 텍스트 투 비디오로 한 번 돌립니다. **32 × 32**, turbo 끔, **40–50** 스텝, 길이는 최종 컷과 같게 맞춥니다. 오디오 latent 프레임 수는 클립 길이로만 정해지고 캔버스 크기와 무관하므로, 이 패스에서도 전체 길이의 오디오가 나오고 비용은 전체 해상도 1스텝 정도입니다. Resolution Selector 연결을 끊고 그것이 물려 있던 `width`와 `height`에 크기를 입력합니다. 이 노드의 **Megapixels** 하한은 **0.1**입니다. 남기는 것은 오디오뿐입니다.
+2. `VAEDecodeAudio`로 디코드하고 `SaveAudioAdvanced`로 파일로 저장합니다.
+3. R2V 워크플로에서 `LoadAudio`를 추가해 그 파일을 불러오고 `ref_audio`에 연결합니다(`audio_vae`는 연결된 채로 둡니다). 프롬프트에서는 `<Audio 1>`으로 인용하고 관계를 밝힙니다. `fully_copy`는 참조한 오디오를 그대로 쓰고, `reference`는 음색만 빌리므로 대사는 직접 씁니다([프롬프트 가이드](/ko/tutorials/video/minimax/minimax-h3-prompt-guide) 참고). 그러면 영상 패스는 4–8 스텝에서 음성을 새로 만드는 대신 이미 정해진 목소리와 타이밍에 맞춥니다.
+
+캔버스 크기는 음질과 시간의 맞바꿈입니다. **32 × 32**는 거의 무료지만 목소리에 기계적인 느낌이 붙을 수 있습니다. **256 × 256**이면 음색이 전체 해상도 렌더에 가까워집니다. 그보다 키우면 이 패스 비용이 최종 렌더만큼 드니, 음성이 최종 품질이어야 할 때만 올리세요. 음성과 가사는 작은 캔버스에서도 대체로 살아남지만, 화면 동작에 붙는 효과음과 스코어는 그렇지 않으므로 마지막 패스에서 만듭니다.
+
 ## 샘플러와 스케줄러
 
 템플릿 **`res_multistep`** + **`simple`**. T2V/I2V/FastH3 서브그래프; R2V·멀티프레임·Fun ControlNet 캔버스 `KSamplerSelect`, `BasicScheduler`.
diff --git a/tutorials/video/minimax/minimax-h3.mdx b/tutorials/video/minimax/minimax-h3.mdx
index d15e44f238..3347bebe96 100644
--- a/tutorials/video/minimax/minimax-h3.mdx
+++ b/tutorials/video/minimax/minimax-h3.mdx
@@ -10,7 +10,7 @@ import UpdateReminder from "/snippets/tutorials/update-reminder.mdx";
 
 Documentation is split across six pages:
 
-- **Overview** (this page): capabilities, checkpoints, workflow index, resolution, clip length, steps, sampler, speedups
+- **Overview** (this page): capabilities, checkpoints, workflow index, resolution, clip length, steps, audio-first pass, sampler, speedups
 - **[Native workflows](/tutorials/video/minimax/minimax-h3-native)**: T2V, I2V, R2V, and native-node techniques
 - **[Multiframe Reference](/tutorials/video/minimax/minimax-h3-multiframe)**: anchor reference frames on the timeline
 - **[Fun ControlNet Union](/tutorials/video/minimax/minimax-h3-fun-controlnet)**: control video or inpainting with a mask
@@ -156,6 +156,16 @@ Guidance by content:
 
 Audio converges later than video in the joint denoise. Speech and timbre need **12+** steps; at **8** steps audio is often the weak link.
 
+## Generating the soundtrack first
+
+Audio and video are denoised together, so short schedules and turbo LoRAs damage the track before the picture. Two passes separate them: settle the audio on a small canvas, then condition the video on it.
+
+1. Run a text-to-video pass of the clip at **32 × 32** with turbo off, **40–50** steps, and the duration the final shot needs. Audio latent frames follow clip length, not canvas size, so this pass renders a full-length track for about the cost of one full-resolution step. Unlink the Resolution Selector and type the size into the `width` and `height` inputs it fed; the selector bottoms out at **0.1** Megapixels. Only the audio is kept.
+2. Decode with `VAEDecodeAudio` and save with `SaveAudioAdvanced`.
+3. In the R2V workflow, add `LoadAudio`, wire it to `ref_audio`, and keep `audio_vae` connected. Cite the track as `<Audio 1>` and mark the relationship: `fully_copy` keeps the reference audio, `reference` borrows only the voice, and then the dialogue is yours to write (see [prompt guide](/tutorials/video/minimax/minimax-h3-prompt-guide)). The video pass then works from a settled voice and timing instead of inventing speech at 4 to 8 steps.
+
+Canvas size trades voice quality against time. **32 × 32** is nearly free, but voices can pick up a robotic edge. **256 × 256** keeps timbre closer to a full-resolution render. Above that the pass costs as much as the render it feeds, so go higher only when the voice has to be final quality. Speech and lyrics usually survive the small canvas; sound effects and scoring tied to on-screen action do not, so render those in the final pass.
+
 ## Sampler and scheduler
 
 Templates use **`res_multistep`** + **`simple`**. T2V/I2V/FastH3 hide these inside the subgraph; R2V, Multiframe, and Fun ControlNet expose `KSamplerSelect` and `BasicScheduler` on canvas.
diff --git a/zh/tutorials/video/minimax/minimax-h3.mdx b/zh/tutorials/video/minimax/minimax-h3.mdx
index b0bea164ff..e3a924a8ee 100644
--- a/zh/tutorials/video/minimax/minimax-h3.mdx
+++ b/zh/tutorials/video/minimax/minimax-h3.mdx
@@ -2,10 +2,10 @@
 title: "ComfyUI MiniMax H3 视频生成指南"
 description: "如何在 ComfyUI 中使用开放权重的 MiniMax H3：文生视频、图生视频、参考生视频工作流，支持原生立体声音频，附提示词编写技巧与 Sage Attention 加速方法。"
 sidebarTitle: "概述"
-translationSourceHash: eb426452
+translationSourceHash: fa293eed
 translationFrom: tutorials/video/minimax/minimax-h3.mdx
 translationBlockHashes:
-  "_intro": 94f54ee8
+  "_intro": 71cca49e
   "Key features": 8ca92280
   "Getting started": eed90954
   "Choosing the checkpoint precision": 4a53fdf9
@@ -13,6 +13,7 @@ translationBlockHashes:
   "Setting the output resolution": 1fda252d
   "Clip length": 2d495d82
   "Step count": 184c83d7
+  "Generating the soundtrack first": 7f97b544
   "Sampler and scheduler": 79c739c3
   "Speeding up generation with Sage Attention": 24cc9a70
   "Speeding up generation with sparse attention": 801c9acf
@@ -25,7 +26,7 @@ import UpdateReminder from "/snippets/zh/tutorials/update-reminder.mdx";
 
 文档分为六个页面：
 
-- **概述**（本页）：能力、检查点、工作流索引、分辨率、片段时长、步数、采样器、加速
+- **概述**（本页）：能力、检查点、工作流索引、分辨率、片段时长、步数、音轨先行、采样器、加速
 - **[原生工作流](/zh/tutorials/video/minimax/minimax-h3-native)**：T2V、I2V、R2V 与原生节点技巧
 - **[多帧参考](/zh/tutorials/video/minimax/minimax-h3-multiframe)**：在时间线上锚定参考帧
 - **[Fun ControlNet Union](/zh/tutorials/video/minimax/minimax-h3-fun-controlnet)**：控制视频或遮罩修补
@@ -171,6 +172,16 @@ Turbo 缩短的是 schedule，不是单步耗时（加载时合并）。4 步 Lo
 
 音频在联合去噪中比画面收敛慢。语音与音色建议 **12+** 步；**8** 步时音频常为弱项。
 
+## 先生成音轨
+
+视频与音频在同一次去噪中生成，步数偏短或用 turbo 时，音轨会先掉质量。分两遍可以拆开：先在极小的画布上定下音频，再让视频以它为条件。
+
+1. 用文生视频跑一遍这个片段：**32 × 32**、关闭 turbo、**40–50** 步，时长与最终视频一致。音频 latent 的帧数只跟片段长度走，与画布大小无关，因此这一遍仍会产出完整长度的音轨，耗时约等于全分辨率的一步。断掉 Resolution Selector 的连线，把尺寸填到它原本接的 `width` 与 `height` 上；该节点的 **Megapixels** 最低只到 **0.1**。这一遍只取音轨。
+2. 用 `VAEDecodeAudio` 解码，再用 `SaveAudioAdvanced` 存成文件。
+3. 在 R2V 工作流里加 `LoadAudio` 载入该文件，接到 `ref_audio`，`audio_vae` 保持连接。提示词里以 `<Audio 1>` 引用它并写明关系：`fully_copy` 沿用这条音轨，`reference` 只借音色，此时台词要自己写（见[提示词指南](/zh/tutorials/video/minimax/minimax-h3-prompt-guide)）。这样视频那遍是对着已定的音色与节奏去对齐，而不是在 4–8 步里现编语音。
+
+画布大小是音质与时间的取舍。**32 × 32** 几乎不花时间，但人声可能带机械感。**256 × 256** 的音色更接近全分辨率渲染；再往上，这一遍的耗时已与它喂给的正片相当，只有音色必须达到成片水准时才值得。语音和歌词通常在小画布上也能成立，与画面动作绑定的音效和配乐则不行，留到最后一遍渲染。
+
 ## 采样器与调度器
 
 模板使用 **`res_multistep`** + **`simple`**。T2V/I2V/FastH3 在子图内；R2V、多帧、Fun ControlNet 在画布暴露 `KSamplerSelect` 与 `BasicScheduler`。

From b2af183f816117d1b29a7d2e1a869bc3dfd6746d Mon Sep 17 00:00:00 2001
From: lin-bot23 <linzhixiong23@gmail.com>
Date: Tue, 27 Oct 2026 18:20:30 +0800
Subject: [PATCH 03/13] docs(minimax): tie the audio-first steps to the T2V
 subgraph and audio markers

---
 es/tutorials/video/minimax/minimax-h3.mdx | 12 ++++++------
 ja/tutorials/video/minimax/minimax-h3.mdx | 12 ++++++------
 ko/tutorials/video/minimax/minimax-h3.mdx | 12 ++++++------
 tutorials/video/minimax/minimax-h3.mdx    |  8 ++++----
 zh/tutorials/video/minimax/minimax-h3.mdx | 12 ++++++------
 5 files changed, 28 insertions(+), 28 deletions(-)

diff --git a/es/tutorials/video/minimax/minimax-h3.mdx b/es/tutorials/video/minimax/minimax-h3.mdx
index 030994fc66..711a7b6469 100644
--- a/es/tutorials/video/minimax/minimax-h3.mdx
+++ b/es/tutorials/video/minimax/minimax-h3.mdx
@@ -2,7 +2,7 @@
 title: "Guía de generación de video con MiniMax H3 en ComfyUI"
 description: "Usa MiniMax H3 de pesos abiertos en ComfyUI: flujos de texto a video, imagen a video y referencia a video con audio estéreo nativo y Sage Attention."
 sidebarTitle: "Descripción general"
-translationSourceHash: fa293eed
+translationSourceHash: d93052eb
 translationFrom: tutorials/video/minimax/minimax-h3.mdx
 translationBlockHashes:
   "_intro": 71cca49e
@@ -13,7 +13,7 @@ translationBlockHashes:
   "Setting the output resolution": 1fda252d
   "Clip length": 2d495d82
   "Step count": 184c83d7
-  "Generating the soundtrack first": 7f97b544
+  "Generating the soundtrack first": a535e97e
   "Sampler and scheduler": 79c739c3
   "Speeding up generation with Sage Attention": 24cc9a70
   "Speeding up generation with sparse attention": 801c9acf
@@ -175,11 +175,11 @@ El audio converge más tarde que el video en el denoise conjunto. El habla y el
 
 El audio y el video se generan en el mismo denoise, así que los schedules cortos y los LoRAs turbo dañan el audio antes que la imagen. Dos pasadas los separan: fija el audio en un lienzo pequeño y después condiciona el video con él.
 
-1. Haz una pasada de texto a video del clip a **32 × 32**, con turbo desactivado, **40–50** pasos y la misma duración que tendrá el plano final. Los fotogramas latentes de audio dependen de la duración del clip, no del tamaño del lienzo, así que la pasada produce una pista completa por un coste cercano al de un solo paso a resolución completa. Desconecta Resolution Selector e introduce el tamaño en los `width` y `height` que alimentaba; su **Megapixels** solo baja hasta **0.1**. De esta pasada solo se conserva el audio.
-2. Decodifícalo con `VAEDecodeAudio` y guárdalo con `SaveAudioAdvanced`.
-3. En el flujo R2V, añade `LoadAudio`, conéctalo a `ref_audio` y deja `audio_vae` conectado. Cita la pista como `<Audio 1>` e indica la relación: `fully_copy` conserva el audio de referencia y `reference` solo toma la voz, en cuyo caso el diálogo lo escribes tú (ver la [guía de prompts](/es/tutorials/video/minimax/minimax-h3-prompt-guide)). La pasada de video trabaja entonces con una voz y un ritmo ya fijados en lugar de inventar habla en 4 a 8 pasos.
+1. Haz una pasada de texto a video del clip a **32 × 32**, turbo desactivado, **40–50** pasos y la misma duración que tendrá el plano final. Los fotogramas latentes de audio dependen de la duración del clip, no del tamaño del lienzo, así que la pasada produce una pista completa por un coste cercano al de un solo paso a resolución completa. Desconecta Resolution Selector e introduce el tamaño en los `width` y `height` que alimentaba; su **Megapixels** solo baja hasta **0.1**. Los pasos se ajustan en el `PrimitiveInt` base dentro del subgrafo T2V (20 por defecto). De esta pasada solo se conserva el audio.
+2. El subgrafo T2V decodifica el audio por dentro y expone solo el video. Ábrelo y conecta su `VAEDecodeAudio` a `SaveAudioAdvanced`.
+3. En el flujo R2V, añade `LoadAudio`, conéctalo a `ref_audio` y deja `audio_vae` conectado. Cita la pista como `<Audio 1>` e indica la relación: `reference` vuelve a generar la frase con esa voz y esa dicción, `fully_copy` toma la referencia como pista final completa y `partially_copy` conserva partes mientras añades efectos o música. La segunda pasada parte entonces de una voz y un ritmo ya fijados en lugar de inventar el audio desde cero. Para obtener el archivo idéntico, conecta `LoadAudio` a `CreateVideo` en vez de confiar en el prompt.
 
-El tamaño del lienzo intercambia calidad de voz por tiempo. **32 × 32** es casi gratis, pero la voz puede adoptar un tono robótico. **256 × 256** acerca el timbre a un render a resolución completa. Por encima, la pasada cuesta tanto como el render al que alimenta, así que sube solo cuando la voz tenga que ser de calidad final. El habla y las letras suelen sobrevivir al lienzo pequeño; los efectos y la música ligados a la acción en pantalla no, así que genéralos en la pasada final.
+El tamaño del lienzo intercambia calidad de voz por tiempo. **32 × 32** es casi gratis, pero la voz puede adoptar un tono robótico. **256 × 256** acerca el timbre a un render a resolución completa. Por encima, la pasada cuesta tanto como el render al que alimenta, así que sube solo cuando la voz tenga que ser de calidad final. El habla y las letras suelen sobrevivir al lienzo pequeño. Los efectos y la música ligados a la acción en pantalla no, así que genéralos en la segunda pasada y deja que `partially_copy` los combine con la pista de la primera.
 
 ## Sampler y scheduler
 
diff --git a/ja/tutorials/video/minimax/minimax-h3.mdx b/ja/tutorials/video/minimax/minimax-h3.mdx
index fb17ae3041..6758c5bd65 100644
--- a/ja/tutorials/video/minimax/minimax-h3.mdx
+++ b/ja/tutorials/video/minimax/minimax-h3.mdx
@@ -2,7 +2,7 @@
 title: "ComfyUI MiniMax H3 動画生成ガイド"
 description: "ComfyUI でオープンウェイトの MiniMax H3 を使う方法: テキストから動画、画像から動画、参照から動画のワークフローとネイティブステレオオーディオ、プロンプト作成のコツ、Sage Attention による高速化。"
 sidebarTitle: "概要"
-translationSourceHash: fa293eed
+translationSourceHash: d93052eb
 translationFrom: tutorials/video/minimax/minimax-h3.mdx
 translationBlockHashes:
   "_intro": 71cca49e
@@ -13,7 +13,7 @@ translationBlockHashes:
   "Setting the output resolution": 1fda252d
   "Clip length": 2d495d82
   "Step count": 184c83d7
-  "Generating the soundtrack first": 7f97b544
+  "Generating the soundtrack first": a535e97e
   "Sampler and scheduler": 79c739c3
   "Speeding up generation with Sage Attention": 24cc9a70
   "Speeding up generation with sparse attention": 801c9acf
@@ -176,11 +176,11 @@ Turbo は schedule を短くするだけで 1 ステップは速くなりませ
 
 映像と音声は同じデノイズで生成されるため、短いスケジュールや turbo では映像より先に音声が崩れます。2 パスに分ければ切り離せます。まず小さなキャンバスで音声を確定し、その音声を条件にして映像を生成します。
 
-1. このクリップをテキストから動画で 1 回生成します。**32 × 32**、turbo オフ、**40–50** ステップ、長さは最終カットと同じにします。音声 latent のフレーム数はクリップ長だけで決まり、キャンバスサイズには依存しないため、このパスでも全長の音声が得られ、コストはフル解像度の 1 ステップ程度です。Resolution Selector のリンクを外し、それが接続していた `width` と `height` にサイズを入力してください。このノードの **Megapixels** は **0.1** が下限です。残すのは音声だけです。
-2. `VAEDecodeAudio` でデコードし、`SaveAudioAdvanced` でファイルに保存します。
-3. R2V ワークフローで `LoadAudio` を追加してそのファイルを読み込み、`ref_audio` に接続します（`audio_vae` は接続したまま）。プロンプトでは `<Audio 1>` として参照し、関係を明記します。`fully_copy` は参照した音声をそのまま使い、`reference` は声質だけを借ります。その場合はセリフを自分で書きます（[プロンプトガイド](/ja/tutorials/video/minimax/minimax-h3-prompt-guide)参照）。これで映像側は、4–8 ステップで音声を作り直すのではなく、確定した声とタイミングに合わせて動きます。
+1. このクリップをテキストから動画で 1 回生成します。**32 × 32**、turbo オフ、**40–50** ステップ、長さは最終カットと同じにします。音声 latent のフレーム数はクリップ長だけで決まり、キャンバスサイズには依存しないため、このパスでも全長の音声が得られ、コストはフル解像度の 1 ステップ程度です。Resolution Selector のリンクを外し、それが接続していた `width` と `height` にサイズを入力してください。このノードの **Megapixels** は **0.1** が下限です。ステップ数は T2V サブグラフ内のベース `PrimitiveInt`（既定 20）で設定します。残すのは音声だけです。
+2. T2V サブグラフは音声を内部でデコードし、映像だけを出力します。サブグラフを開き、その `VAEDecodeAudio` を `SaveAudioAdvanced` に接続してください。
+3. R2V ワークフローで `LoadAudio` を追加してそのファイルを読み込み、`ref_audio` に接続します（`audio_vae` は接続したまま）。プロンプトでは `<Audio 1>` として参照し、関係を明記します。`reference` はその声と話し方でセリフを生成し直し、`fully_copy` は参照音声を最終トラックそのものとして扱い、`partially_copy` は一部を残して効果音や音楽を重ねます。これで 2 パス目は確定した声とタイミングから始まり、ゼロから音声を作ることはありません。ファイルと同一の音声にするなら、プロンプトに頼らず `LoadAudio` を `CreateVideo` に直接接続します。
 
-キャンバスサイズは音質と時間のトレードオフです。**32 × 32** はほぼ無コストですが、声に機械的な響きが出ることがあります。**256 × 256** なら音色はフル解像度のレンダーに近づきます。それ以上にすると、このパスのコストは供給先のレンダーと同程度になるため、音声を最終品質にしたいときだけ上げてください。セリフや歌詞は小さなキャンバスでも大抵成立しますが、画面の動きに紐づく効果音やスコアは成立しないため、最終パスで生成します。
+キャンバスサイズは音質と時間のトレードオフです。**32 × 32** はほぼ無コストですが、声に機械的な響きが出ることがあります。**256 × 256** なら音色はフル解像度のレンダーに近づきます。それ以上にすると、このパスのコストは供給先のレンダーと同程度になるため、音声を最終品質にしたいときだけ上げてください。セリフや歌詞は小さなキャンバスでも大抵成立します。画面の動きに紐づく効果音やスコアは成立しないため、2 パス目で生成し、`partially_copy` で 1 パス目の音声と合わせます。
 
 ## サンプラーとスケジューラ
 
diff --git a/ko/tutorials/video/minimax/minimax-h3.mdx b/ko/tutorials/video/minimax/minimax-h3.mdx
index 0d4da36fee..f1ea4920e1 100644
--- a/ko/tutorials/video/minimax/minimax-h3.mdx
+++ b/ko/tutorials/video/minimax/minimax-h3.mdx
@@ -2,7 +2,7 @@
 title: "ComfyUI MiniMax H3 비디오 생성 가이드"
 description: "ComfyUI에서 오픈 가중치 MiniMax H3를 사용하는 방법: 텍스트 기반, 이미지 기반, 레퍼런스 기반 비디오 워크플로와 네이티브 스테레오 오디오, 프롬프트 작성 팁, Sage Attention 가속화."
 sidebarTitle: "개요"
-translationSourceHash: fa293eed
+translationSourceHash: d93052eb
 translationFrom: tutorials/video/minimax/minimax-h3.mdx
 translationBlockHashes:
   "_intro": 71cca49e
@@ -13,7 +13,7 @@ translationBlockHashes:
   "Setting the output resolution": 1fda252d
   "Clip length": 2d495d82
   "Step count": 184c83d7
-  "Generating the soundtrack first": 7f97b544
+  "Generating the soundtrack first": a535e97e
   "Sampler and scheduler": 79c739c3
   "Speeding up generation with Sage Attention": 24cc9a70
   "Speeding up generation with sparse attention": 801c9acf
@@ -176,11 +176,11 @@ Turbo는 schedule만 단축, 스텝당 시간 동일(로드 시 병합). 4스텝
 
 영상과 오디오는 같은 디노이즈에서 함께 생성되므로, 스텝이 짧거나 turbo를 쓰면 영상보다 오디오가 먼저 무너집니다. 두 번의 패스로 분리할 수 있습니다. 작은 캔버스에서 오디오를 먼저 확정하고, 그 오디오를 조건으로 영상을 생성합니다.
 
-1. 이 클립을 텍스트 투 비디오로 한 번 돌립니다. **32 × 32**, turbo 끔, **40–50** 스텝, 길이는 최종 컷과 같게 맞춥니다. 오디오 latent 프레임 수는 클립 길이로만 정해지고 캔버스 크기와 무관하므로, 이 패스에서도 전체 길이의 오디오가 나오고 비용은 전체 해상도 1스텝 정도입니다. Resolution Selector 연결을 끊고 그것이 물려 있던 `width`와 `height`에 크기를 입력합니다. 이 노드의 **Megapixels** 하한은 **0.1**입니다. 남기는 것은 오디오뿐입니다.
-2. `VAEDecodeAudio`로 디코드하고 `SaveAudioAdvanced`로 파일로 저장합니다.
-3. R2V 워크플로에서 `LoadAudio`를 추가해 그 파일을 불러오고 `ref_audio`에 연결합니다(`audio_vae`는 연결된 채로 둡니다). 프롬프트에서는 `<Audio 1>`으로 인용하고 관계를 밝힙니다. `fully_copy`는 참조한 오디오를 그대로 쓰고, `reference`는 음색만 빌리므로 대사는 직접 씁니다([프롬프트 가이드](/ko/tutorials/video/minimax/minimax-h3-prompt-guide) 참고). 그러면 영상 패스는 4–8 스텝에서 음성을 새로 만드는 대신 이미 정해진 목소리와 타이밍에 맞춥니다.
+1. 이 클립을 텍스트 투 비디오로 한 번 돌립니다. **32 × 32**, turbo 끔, **40–50** 스텝, 길이는 최종 컷과 같게 맞춥니다. 오디오 latent 프레임 수는 클립 길이로만 정해지고 캔버스 크기와 무관하므로, 이 패스에서도 전체 길이의 오디오가 나오고 비용은 전체 해상도 1스텝 정도입니다. Resolution Selector 연결을 끊고 그것이 물려 있던 `width`와 `height`에 크기를 입력합니다. 이 노드의 **Megapixels** 하한은 **0.1**입니다. 스텝 수는 T2V 서브그래프 안의 기본 `PrimitiveInt`(기본값 20)에서 설정합니다. 남기는 것은 오디오뿐입니다.
+2. T2V 서브그래프는 오디오를 내부에서 디코드하고 영상만 출력합니다. 서브그래프를 열어 그 `VAEDecodeAudio`를 `SaveAudioAdvanced`에 연결하세요.
+3. R2V 워크플로에서 `LoadAudio`를 추가해 그 파일을 불러오고 `ref_audio`에 연결합니다(`audio_vae`는 연결된 채로 둡니다). 프롬프트에서는 `<Audio 1>`으로 인용하고 관계를 밝힙니다. `reference`는 그 목소리와 말투로 대사를 다시 생성하고, `fully_copy`는 참조 오디오를 최종 트랙 자체로 취급하며, `partially_copy`는 일부를 유지한 채 효과음이나 음악을 얹습니다. 그러면 두 번째 패스는 이미 정해진 목소리와 타이밍에서 시작하고, 처음부터 음성을 만들지 않습니다. 파일과 똑같은 오디오를 쓰려면 프롬프트에 기대지 말고 `LoadAudio`를 `CreateVideo`에 바로 연결합니다.
 
-캔버스 크기는 음질과 시간의 맞바꿈입니다. **32 × 32**는 거의 무료지만 목소리에 기계적인 느낌이 붙을 수 있습니다. **256 × 256**이면 음색이 전체 해상도 렌더에 가까워집니다. 그보다 키우면 이 패스 비용이 최종 렌더만큼 드니, 음성이 최종 품질이어야 할 때만 올리세요. 음성과 가사는 작은 캔버스에서도 대체로 살아남지만, 화면 동작에 붙는 효과음과 스코어는 그렇지 않으므로 마지막 패스에서 만듭니다.
+캔버스 크기는 음질과 시간의 맞바꿈입니다. **32 × 32**는 거의 무료지만 목소리에 기계적인 느낌이 붙을 수 있습니다. **256 × 256**이면 음색이 전체 해상도 렌더에 가까워집니다. 그보다 키우면 이 패스 비용이 최종 렌더만큼 드니, 음성이 최종 품질이어야 할 때만 올리세요. 음성과 가사는 작은 캔버스에서도 대체로 살아남습니다. 화면 동작에 붙는 효과음과 스코어는 그렇지 않으므로 두 번째 패스에서 만들고, `partially_copy`로 첫 패스 오디오와 합칩니다.
 
 ## 샘플러와 스케줄러
 
diff --git a/tutorials/video/minimax/minimax-h3.mdx b/tutorials/video/minimax/minimax-h3.mdx
index 3347bebe96..9fa3827f21 100644
--- a/tutorials/video/minimax/minimax-h3.mdx
+++ b/tutorials/video/minimax/minimax-h3.mdx
@@ -160,11 +160,11 @@ Audio converges later than video in the joint denoise. Speech and timbre need **
 
 Audio and video are denoised together, so short schedules and turbo LoRAs damage the track before the picture. Two passes separate them: settle the audio on a small canvas, then condition the video on it.
 
-1. Run a text-to-video pass of the clip at **32 × 32** with turbo off, **40–50** steps, and the duration the final shot needs. Audio latent frames follow clip length, not canvas size, so this pass renders a full-length track for about the cost of one full-resolution step. Unlink the Resolution Selector and type the size into the `width` and `height` inputs it fed; the selector bottoms out at **0.1** Megapixels. Only the audio is kept.
-2. Decode with `VAEDecodeAudio` and save with `SaveAudioAdvanced`.
-3. In the R2V workflow, add `LoadAudio`, wire it to `ref_audio`, and keep `audio_vae` connected. Cite the track as `<Audio 1>` and mark the relationship: `fully_copy` keeps the reference audio, `reference` borrows only the voice, and then the dialogue is yours to write (see [prompt guide](/tutorials/video/minimax/minimax-h3-prompt-guide)). The video pass then works from a settled voice and timing instead of inventing speech at 4 to 8 steps.
+1. Run a text-to-video pass of the clip at **32 × 32**, turbo off, **40–50** steps, and the duration the final shot needs. Audio latent frames follow clip length, not canvas size, so this pass renders a full-length track for about the cost of one full-resolution step. Unlink the Resolution Selector and type the size into the `width` and `height` inputs it fed; the selector bottoms out at **0.1** Megapixels. Steps live on the base `PrimitiveInt` inside the T2V subgraph (20 by default). Only the audio is kept.
+2. The T2V subgraph decodes audio inside and exposes only the video. Open the subgraph and connect its `VAEDecodeAudio` to `SaveAudioAdvanced`.
+3. In the R2V workflow, add `LoadAudio`, wire it to `ref_audio`, and keep `audio_vae` connected. Cite the track as `<Audio 1>` and state the relationship: `reference` regenerates the line in that voice and delivery, `fully_copy` treats the reference as the complete final track, and `partially_copy` keeps parts of it while you add effects or music. The second pass then starts from a settled voice and timing instead of inventing speech from nothing. For the identical file, wire `LoadAudio` into `CreateVideo` instead of relying on the prompt.
 
-Canvas size trades voice quality against time. **32 × 32** is nearly free, but voices can pick up a robotic edge. **256 × 256** keeps timbre closer to a full-resolution render. Above that the pass costs as much as the render it feeds, so go higher only when the voice has to be final quality. Speech and lyrics usually survive the small canvas; sound effects and scoring tied to on-screen action do not, so render those in the final pass.
+Canvas size trades voice quality against time. **32 × 32** is nearly free, but voices can pick up a robotic edge. **256 × 256** keeps timbre closer to a full-resolution render. Above that the pass costs as much as the render it feeds, so go higher only when the voice has to be final quality. Speech and lyrics usually survive the small canvas. Sound effects and scoring tied to on-screen action do not, so render those in the second pass and let `partially_copy` combine them with the first track.
 
 ## Sampler and scheduler
 
diff --git a/zh/tutorials/video/minimax/minimax-h3.mdx b/zh/tutorials/video/minimax/minimax-h3.mdx
index e3a924a8ee..cba6d43df2 100644
--- a/zh/tutorials/video/minimax/minimax-h3.mdx
+++ b/zh/tutorials/video/minimax/minimax-h3.mdx
@@ -2,7 +2,7 @@
 title: "ComfyUI MiniMax H3 视频生成指南"
 description: "如何在 ComfyUI 中使用开放权重的 MiniMax H3：文生视频、图生视频、参考生视频工作流，支持原生立体声音频，附提示词编写技巧与 Sage Attention 加速方法。"
 sidebarTitle: "概述"
-translationSourceHash: fa293eed
+translationSourceHash: d93052eb
 translationFrom: tutorials/video/minimax/minimax-h3.mdx
 translationBlockHashes:
   "_intro": 71cca49e
@@ -13,7 +13,7 @@ translationBlockHashes:
   "Setting the output resolution": 1fda252d
   "Clip length": 2d495d82
   "Step count": 184c83d7
-  "Generating the soundtrack first": 7f97b544
+  "Generating the soundtrack first": a535e97e
   "Sampler and scheduler": 79c739c3
   "Speeding up generation with Sage Attention": 24cc9a70
   "Speeding up generation with sparse attention": 801c9acf
@@ -176,11 +176,11 @@ Turbo 缩短的是 schedule，不是单步耗时（加载时合并）。4 步 Lo
 
 视频与音频在同一次去噪中生成，步数偏短或用 turbo 时，音轨会先掉质量。分两遍可以拆开：先在极小的画布上定下音频，再让视频以它为条件。
 
-1. 用文生视频跑一遍这个片段：**32 × 32**、关闭 turbo、**40–50** 步，时长与最终视频一致。音频 latent 的帧数只跟片段长度走，与画布大小无关，因此这一遍仍会产出完整长度的音轨，耗时约等于全分辨率的一步。断掉 Resolution Selector 的连线，把尺寸填到它原本接的 `width` 与 `height` 上；该节点的 **Megapixels** 最低只到 **0.1**。这一遍只取音轨。
-2. 用 `VAEDecodeAudio` 解码，再用 `SaveAudioAdvanced` 存成文件。
-3. 在 R2V 工作流里加 `LoadAudio` 载入该文件，接到 `ref_audio`，`audio_vae` 保持连接。提示词里以 `<Audio 1>` 引用它并写明关系：`fully_copy` 沿用这条音轨，`reference` 只借音色，此时台词要自己写（见[提示词指南](/zh/tutorials/video/minimax/minimax-h3-prompt-guide)）。这样视频那遍是对着已定的音色与节奏去对齐，而不是在 4–8 步里现编语音。
+1. 用文生视频跑一遍这个片段：**32 × 32**、关闭 turbo、**40–50** 步，时长与最终视频一致。音频 latent 的帧数只跟片段长度走，与画布大小无关，因此这一遍仍会产出完整长度的音轨，耗时约等于全分辨率的一步。断掉 Resolution Selector 的连线，把尺寸填到它原本接的 `width` 与 `height` 上；该节点的 **Megapixels** 最低只到 **0.1**。步数在 T2V 子图内部的基准 `PrimitiveInt` 上设置（默认 20）。这一遍只取音轨。
+2. T2V 子图在内部解码音频，只把视频暴露出来。进入子图，把它的 `VAEDecodeAudio` 接到 `SaveAudioAdvanced`。
+3. 在 R2V 工作流里加 `LoadAudio` 载入该文件，接到 `ref_audio`，`audio_vae` 保持连接。提示词里以 `<Audio 1>` 引用并写明关系：`reference` 用这条音轨的声线与语气重新生成台词，`fully_copy` 把它当作完整的最终音轨，`partially_copy` 保留其中一部分、再叠加音效或配乐。这样第二遍是接着已定的音色与节奏走，而不是从零现编语音。要得到与原文件完全一致的音轨，把 `LoadAudio` 直接接进 `CreateVideo`，不要依赖提示词。
 
-画布大小是音质与时间的取舍。**32 × 32** 几乎不花时间，但人声可能带机械感。**256 × 256** 的音色更接近全分辨率渲染；再往上，这一遍的耗时已与它喂给的正片相当，只有音色必须达到成片水准时才值得。语音和歌词通常在小画布上也能成立，与画面动作绑定的音效和配乐则不行，留到最后一遍渲染。
+画布大小是音质与时间的取舍。**32 × 32** 几乎不花时间，但人声可能带机械感。**256 × 256** 的音色更接近全分辨率渲染；再往上，这一遍的耗时已与它喂给的正片相当，只有音色必须达到成片水准时才值得。语音和歌词通常在小画布上也能成立。与画面动作绑定的音效和配乐则不行，留到第二遍渲染，并用 `partially_copy` 把两段合成。
 
 ## 采样器与调度器
 

From 564e9fc2d7a5acf10b812554bf5e11901347c33f Mon Sep 17 00:00:00 2001
From: lin-bot23 <linzhixiong23@gmail.com>
Date: Wed, 28 Oct 2026 18:08:06 +0800
Subject: [PATCH 04/13] docs(flux): explain the FLUX 3 Klein bounding box
 layout path

---
 es/tutorials/flux/flux-3-klein.mdx | 14 ++++++++++----
 ja/tutorials/flux/flux-3-klein.mdx | 15 ++++++++++-----
 ko/tutorials/flux/flux-3-klein.mdx | 15 ++++++++++-----
 tutorials/flux/flux-3-klein.mdx    |  9 +++++++--
 zh/tutorials/flux/flux-3-klein.mdx | 15 ++++++++++-----
 5 files changed, 47 insertions(+), 21 deletions(-)

diff --git a/es/tutorials/flux/flux-3-klein.mdx b/es/tutorials/flux/flux-3-klein.mdx
index ff263a9586..8135ca5890 100644
--- a/es/tutorials/flux/flux-3-klein.mdx
+++ b/es/tutorials/flux/flux-3-klein.mdx
@@ -2,13 +2,14 @@
 title: "FLUX 3 Image Klein en ComfyUI: texto a imagen y edición"
 description: "Ejecuta FLUX.3 [klein] en local en ComfyUI: texto a imagen con pesos abiertos y edición por instrucciones, con un checkpoint destilado para 4–16 pasos."
 sidebarTitle: "FLUX 3 Klein"
-translationSourceHash: 5c435b1b
+translationSourceHash: 1011d294
 translationFrom: tutorials/flux/flux-3-klein.mdx
 translationBlockHashes:
   "_intro": 3e9c5a21
   "About FLUX.3 [klein]": 1d7659ce
   "FLUX 3 Image Klein Workflows": d0941358
-  "Prompting and reference tips": 43ad4a36
+  "FLUX 3 Image Klein Model Downloads": 2972cb13
+  "Prompting and reference tips": 9f809dad
 ---
 import UpdateReminder from "/snippets/es/tutorials/update-reminder.mdx"
 
@@ -174,7 +175,7 @@ Las plantillas cargan las compilaciones int8 para mantener pequeño el consumo d
     Codificador de texto para ambos checkpoints. Cárgalo con `CLIPLoader` de tipo `flux3`.
   </Card>
 <Card title="qwen3.5_9b_flux3_prompt_upsampler_w4a8.safetensors" icon="download" href="https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy">
-    Upsampler de prompt que usan las plantillas de texto a imagen.
+    Upsampler de prompt para prompts cortos. Las cuatro plantillas lo cargan.
   </Card>
 <Card title="flux2-vae.safetensors" icon="download" href="https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy">
     VAE para FLUX 3 Klein. El VAE de FLUX.2.
@@ -202,4 +203,9 @@ Las plantillas cargan las compilaciones int8 para mantener pequeño el consumo d
 - **Las referencias se añaden a ambos conditioning.** Conecta imágenes adicionales a `Set Reference Multi Latent` en orden, llámalas `<ref_image_0>`, `<ref_image_1>` y así sucesivamente, y añade cada una tanto al prompt positivo como al negativo, igual que hace el pipeline de referencia.
 - **El tamaño de la referencia es un control de costo.** Las plantillas redimensionan cada referencia a aproximadamente 1 megapíxel y el latente de salida toma ese tamaño. Mantener las referencias en su tamaño nativo, que es lo que hace el pipeline propio de Black Forest Labs hasta unos 16 megapíxeles, conserva más detalle en la edición a un costo mayor.
 - **La cantidad de pasos depende del checkpoint.** El checkpoint completo se ejecuta con 50 pasos en las plantillas, con orientación 3.5 en texto a imagen y 1 en edición de imagen. El checkpoint destilado por pasos se ejecuta con 16 pasos y orientación fija en 1, y es el que debes usar cuando quieras una pasada rápida.
-- **Las plantillas incluyen un nodo `Create Bounding Boxes`.** Dibuja cajas sobre la imagen fuente para planificar ediciones regionales, y su lista de cajas se muestra en el nodo `Preview Any` para que puedas integrar las regiones en el prompt.
+- **Las cajas llegan al modelo a través de `Flux3LayoutPrompt`.** `Create Bounding Boxes` entrega las cajas en su salida `elements`, y `Flux3LayoutPrompt` las añade al prompt como una tabla de diseño JSON con un `id`, un `bbox` y un `desc` por caja. Las cajas se numeran `obj_1`, `obj_2` y así sucesivamente en el orden en que las dibujas, así que dibuja primero el elemento de fondo. Cita una caja en el prompt como `<obj_1>` para vincular las palabras a esa caja. Si no hay cajas dibujadas, el nodo devuelve el prompt sin cambios.
+- **Las plantillas de edición de imagen ya usan las cajas.** Llevan `Flux3LayoutPrompt` a través del upsampler de prompt y un switch hasta el codificador de texto del sampler, así que la lista de cajas llega al modelo sin recablear nada. Escribe la frase de edición en el campo de prompt del propio nodo de diseño y deja `<obj_1>` donde corresponde la caja.
+- **Las plantillas de texto a imagen dejan la salida de diseño sin conectar.** Solo llega a `Preview Any`. Escribe el prompt en `Flux3LayoutPrompt` y arrastra su salida `prompt` hasta la entrada de prompt del subgrafo. Desactiva `enable_prompt_upsampler` antes: si no, el upsampler Qwen3.5 reescribe ese texto, tabla de diseño incluida, antes de muestrear.
+- **En texto a imagen, cubre todo el lienzo con una sola caja.** Escribe la escena en la descripción de esa caja. El modelo recibe entonces una fila de colocación para todo el encuadre en lugar de texto sin información de diseño.
+- **Mantén el lienzo de las cajas con la relación de aspecto de la salida.** Las coordenadas se guardan como fracciones del `width` y el `height` de `Create Bounding Boxes`, así que un lienzo con una relación de aspecto distinta a la del latente de salida estira las regiones. Cualquier tamaño de lienzo cuadra si la proporción coincide. Las plantillas de edición dibujan sobre un lienzo de 896 x 1152 y su latente es de 1024 x 1024, así que cambia el tamaño del latente al del lienzo cuando necesites regiones exactas.
+- **Los colores van en la descripción de la caja.** La paleta del editor se ignora, así que nombra los colores en la descripción.
diff --git a/ja/tutorials/flux/flux-3-klein.mdx b/ja/tutorials/flux/flux-3-klein.mdx
index 64a541f423..aa2d32b35c 100644
--- a/ja/tutorials/flux/flux-3-klein.mdx
+++ b/ja/tutorials/flux/flux-3-klein.mdx
@@ -2,14 +2,14 @@
 title: "FLUX 3 Image Klein ComfyUI ガイド"
 description: "ComfyUI で FLUX.3 [klein] をローカル実行: オープンウェイトのテキストから画像生成と指示による画像編集。4〜16 ステップの蒸留版にも対応。"
 sidebarTitle: "FLUX 3 Klein"
-translationSourceHash: 7cedfbb4
+translationSourceHash: 1011d294
 translationFrom: tutorials/flux/flux-3-klein.mdx
 translationBlockHashes:
   "_intro": 3e9c5a21
   "About FLUX.3 [klein]": 1d7659ce
   "FLUX 3 Image Klein Workflows": d0941358
-  "FLUX 3 Image Klein Model Downloads": 0170574c
-  "Prompting and reference tips": 43ad4a36
+  "FLUX 3 Image Klein Model Downloads": 2972cb13
+  "Prompting and reference tips": 9f809dad
 ---
 import UpdateReminder from "/snippets/ja/tutorials/update-reminder.mdx"
 
@@ -175,7 +175,7 @@ FLUX 3 Klein の重みは、ライセンスへの同意が必要なリポジト
     両方の checkpoint で使うテキストエンコーダー。`CLIPLoader` のタイプ `flux3` で読み込みます。
   </Card>
 <Card title="qwen3.5_9b_flux3_prompt_upsampler_w4a8.safetensors" icon="download" href="https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy">
-    テキストから画像へのテンプレートで使うプロンプトアップサンプラー。
+    短いプロンプト用のプロンプトアップサンプラー。4 つのテンプレートすべてが読み込みます。
   </Card>
 <Card title="flux2-vae.safetensors" icon="download" href="https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy">
     FLUX 3 Klein の VAE。FLUX.2 VAE。
@@ -203,4 +203,9 @@ FLUX 3 Klein の重みは、ライセンスへの同意が必要なリポジト
 - **参照は両方の conditioning に追加されます。** 追加の画像を順番に `Set Reference Multi Latent` に接続し、`<ref_image_0>`、`<ref_image_1>` などとして参照し、参照パイプラインが行うのと同じように、それぞれをポジティブプロンプトとネガティブプロンプトの両方に追加します。
 - **参照のサイズはコストの調整ダイヤルです。** テンプレートは各参照を約1メガピクセルにリサイズし、出力潜在はそのサイズになります。Black Forest Labs 自身のパイプラインが最大約16メガピクセルまで対応しているように、参照をネイティブサイズのままにしておくと、より高いコストと引き換えに編集でより多くのディテールを保てます。
 - **ステップ数はチェックポイントによって異なります。** テンプレートでは、フルのチェックポイントは50ステップで実行され、テキストから画像ではガイダンス3.5、画像編集では1になります。ステップ蒸留版のチェックポイントは16ステップ、ガイダンスは1に固定で実行され、高速に処理したいときに使うのはこちらです。
-- **テンプレートには `Create Bounding Boxes` ノードが付属しています。** これは領域ごとの編集を計画するためにソース画像の上にボックスを描画するもので、そのボックスのリストは `Preview Any` ノードに表示されるので、領域をプロンプトに織り込むことができます。
+- **ボックスは `Flux3LayoutPrompt` 経由でモデルに渡ります。** `Create Bounding Boxes` はボックスを `elements` 出力に出し、`Flux3LayoutPrompt` がそれを `id`、`bbox`、`desc` を持つ JSON レイアウト表としてプロンプトの末尾に追加します。ボックスは描いた順に `obj_1`、`obj_2` と番号が振られるので、背景の要素を先に描いてください。プロンプト内で `<obj_1>` として参照すると、その語がそのボックスに紐づきます。ボックスを描いていない場合、このノードはプロンプトをそのまま返します。
+- **画像編集テンプレートはすでにボックスを使っています。** `Flux3LayoutPrompt` をプロンプトアップサンプラとスイッチ経由でサンプラーのテキストエンコードに接続しているため、配線を変えなくてもボックス一覧がモデルに届きます。編集の文はこのレイアウトノード自身のプロンプト欄に書き、ボックスに対応する位置に `<obj_1>` を残します。
+- **テキストから画像のテンプレートはレイアウト出力が未接続です。** `Preview Any` にしかつながっていません。`Flux3LayoutPrompt` にプロンプトを書き、その `prompt` 出力をサブグラフのプロンプト入力にドラッグします。先に `enable_prompt_upsampler` を切ってください。切らないと Qwen3.5 のアップサンプラがサンプリング前にこのテキストを書き換え、レイアウト表も一緒に変わります。
+- **テキストから画像では、キャンバス全体を覆う 1 つのボックスを使います。** シーンの説明はそのボックスの説明欄に書きます。モデルはレイアウト情報のないテキストではなく、フレーム全体に対する配置行を受け取ります。
+- **ボックスのキャンバスは出力と同じアスペクト比にします。** ボックス座標は `Create Bounding Boxes` の `width` と `height` に対する比率で保存されるため、キャンバスのアスペクト比が出力潜在と異なると領域が引き伸ばされます。比率が同じならキャンバスのピクセル数はいくつでも合います。画像編集テンプレートのキャンバスは 896 x 1152、潜在は 1024 x 1024 なので、正確な領域が必要なときは潜在をキャンバスと同じ比率に変更してください。
+- **色はボックスの説明欄に書きます。** エディタのパレットは無視されるため、色は説明に書いてください。
diff --git a/ko/tutorials/flux/flux-3-klein.mdx b/ko/tutorials/flux/flux-3-klein.mdx
index 74bde405bb..45f7105162 100644
--- a/ko/tutorials/flux/flux-3-klein.mdx
+++ b/ko/tutorials/flux/flux-3-klein.mdx
@@ -2,14 +2,14 @@
 title: "FLUX 3 Image Klein ComfyUI 가이드"
 description: "ComfyUI에서 FLUX.3 [klein]을 로컬로 실행: 오픈 웨이트 텍스트 기반 이미지 생성과 지시 기반 이미지 편집, 4~16스텝 증류판 포함."
 sidebarTitle: "FLUX 3 Klein"
-translationSourceHash: 7cedfbb4
+translationSourceHash: 1011d294
 translationFrom: tutorials/flux/flux-3-klein.mdx
 translationBlockHashes:
   "_intro": 3e9c5a21
   "About FLUX.3 [klein]": 1d7659ce
   "FLUX 3 Image Klein Workflows": d0941358
-  "FLUX 3 Image Klein Model Downloads": 0170574c
-  "Prompting and reference tips": 43ad4a36
+  "FLUX 3 Image Klein Model Downloads": 2972cb13
+  "Prompting and reference tips": 9f809dad
 ---
 import UpdateReminder from "/snippets/ko/tutorials/update-reminder.mdx"
 
@@ -175,7 +175,7 @@ FLUX 3 Klein 가중치는 라이선스 동의가 필요한 저장소 [black-fore
     두 checkpoint가 함께 사용하는 텍스트 인코더입니다. `CLIPLoader` 유형 `flux3`으로 불러오세요.
   </Card>
 <Card title="qwen3.5_9b_flux3_prompt_upsampler_w4a8.safetensors" icon="download" href="https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy">
-    텍스트 기반 이미지 생성 템플릿에서 사용하는 프롬프트 업샘플러입니다.
+    짧은 프롬프트를 위한 프롬프트 업샘플러입니다. 네 템플릿 모두 불러옵니다.
   </Card>
 <Card title="flux2-vae.safetensors" icon="download" href="https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy">
     FLUX 3 Klein의 VAE, 즉 FLUX.2 VAE.
@@ -203,4 +203,9 @@ FLUX 3 Klein 가중치는 라이선스 동의가 필요한 저장소 [black-fore
 - **레퍼런스는 두 conditioning 모두에 추가됩니다.** 추가 이미지를 `Set Reference Multi Latent`에 순서대로 연결하고, `<ref_image_0>`, `<ref_image_1>` 등으로 지칭한 다음, 레퍼런스 pipeline이 하는 방식대로 각각을 positive 프롬프트와 네거티브 프롬프트 양쪽에 추가하세요.
 - **레퍼런스 크기는 비용을 조절하는 다이얼입니다.** 템플릿은 각 레퍼런스의 크기를 약 1메가픽셀로 조정하며, 출력 잠재 데이터는 그 크기를 갖습니다. Black Forest Labs 자체 pipeline이 약 16메가픽셀까지 유지하는 것처럼 레퍼런스를 원본 크기로 유지하면 비용이 더 높아지는 대신 편집 결과에 더 많은 디테일이 남습니다.
 - **스텝 수는 checkpoint에 따라 다릅니다.** 전체 checkpoint는 템플릿에서 50 스텝으로 실행되며, 텍스트 기반 생성에서는 가이드 3.5, 이미지 편집에서는 1입니다. 스텝 증류(step-distilled) checkpoint는 16 스텝에 가이드가 1로 고정되어 실행되며, 빠르게 처리하고 싶을 때 사용하는 것이 바로 이 checkpoint입니다.
-- **템플릿에는 `Create Bounding Boxes` 노드가 포함되어 있습니다.** 이 노드는 소스 이미지 위에 박스를 그려 영역별 편집을 계획할 수 있게 해주며, 박스 목록은 `Preview Any` 노드에 표시되므로 해당 영역을 프롬프트에 반영할 수 있습니다.
+- **박스는 `Flux3LayoutPrompt`를 거쳐 모델에 전달됩니다.** `Create Bounding Boxes`는 박스를 `elements` 출력으로 내보내고, `Flux3LayoutPrompt`는 이를 `id`, `bbox`, `desc`를 담은 JSON 레이아웃 표로 프롬프트 뒤에 덧붙입니다. 박스는 그린 순서대로 `obj_1`, `obj_2`로 번호가 붙으므로 배경 요소를 먼저 그리세요. 프롬프트에서 `<obj_1>`로 인용하면 해당 단어가 그 박스에 연결됩니다. 박스를 그리지 않으면 이 노드는 프롬프트를 그대로 반환합니다.
+- **이미지 편집 템플릿은 이미 박스를 사용합니다.** `Flux3LayoutPrompt`를 프롬프트 업샘플러와 스위치를 거쳐 샘플러의 텍스트 인코드로 보내므로, 배선을 바꾸지 않아도 박스 목록이 모델에 전달됩니다. 편집 문장은 이 레이아웃 노드 자체의 프롬프트 칸에 쓰고, 박스에 해당하는 자리에 `<obj_1>`을 남겨 두세요.
+- **텍스트에서 이미지 템플릿은 레이아웃 출력이 연결되어 있지 않습니다.** `Preview Any`에만 연결됩니다. `Flux3LayoutPrompt`에 프롬프트를 쓰고 그 `prompt` 출력을 서브그래프의 프롬프트 입력으로 끌어다 놓으세요. 먼저 `enable_prompt_upsampler`를 끄세요. 끄지 않으면 Qwen3.5 업샘플러가 샘플링 전에 이 텍스트를 다시 쓰면서 레이아웃 표도 함께 바뀝니다.
+- **텍스트에서 이미지 생성 시 캔버스 전체를 덮는 박스 하나를 사용하세요.** 장면 설명은 그 박스의 설명에 적습니다. 그러면 모델은 레이아웃 정보가 없는 텍스트 대신 프레임 전체에 대한 배치 행을 받습니다.
+- **박스 캔버스는 출력과 같은 가로세로 비율로 맞추세요.** 박스 좌표는 `Create Bounding Boxes`의 `width`와 `height`에 대한 비율로 저장되므로, 캔버스 비율이 출력 잠재 데이터와 다르면 영역이 늘어납니다. 비율만 같으면 캔버스 픽셀 크기는 상관없습니다. 이미지 편집 템플릿의 캔버스는 896 x 1152이고 잠재 데이터는 1024 x 1024이므로, 정확한 영역이 필요하면 잠재 크기를 캔버스 비율에 맞게 바꾸세요.
+- **색상은 박스 설명에 적습니다.** 편집기의 팔레트는 무시되므로 색을 설명에 적으세요.
diff --git a/tutorials/flux/flux-3-klein.mdx b/tutorials/flux/flux-3-klein.mdx
index b052d18672..d080b6fbbf 100644
--- a/tutorials/flux/flux-3-klein.mdx
+++ b/tutorials/flux/flux-3-klein.mdx
@@ -168,7 +168,7 @@ The templates load the int8 builds to keep the memory footprint small. With more
     Text encoder for both checkpoints. Load it with `CLIPLoader` type `flux3`.
   </Card>
 <Card title="qwen3.5_9b_flux3_prompt_upsampler_w4a8.safetensors" icon="download" href="https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy">
-    Prompt upsampler used by the text to image templates.
+    Prompt upsampler for short prompts. All four templates load it.
   </Card>
 <Card title="flux2-vae.safetensors" icon="download" href="https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy">
     VAE for FLUX 3 Klein. The FLUX.2 VAE.
@@ -196,4 +196,9 @@ The templates load the int8 builds to keep the memory footprint small. With more
 - **References are added to both conditionings.** Connect extra images to `Set Reference Multi Latent` in order, address them as `<ref_image_0>`, `<ref_image_1>`, and so on, and add each one to both the positive and the negative prompt, the way the reference pipeline does.
 - **Reference size is a cost dial.** The templates resize each reference to roughly 1 megapixel and the output latent takes that size. Keeping references at their native size, which Black Forest Labs' own pipeline does up to about 16 megapixels, keeps more detail in the edit at a higher cost.
 - **Step count depends on the checkpoint.** The full checkpoint runs at 50 steps in the templates, with guidance 3.5 on text to image and 1 on image edit. The step-distilled checkpoint runs at 16 steps with guidance fixed at 1, and it is the one to use when you want a fast pass.
-- **The templates ship a `Create Bounding Boxes` node.** It draws boxes over the source image for planning regional edits, and its list of boxes is shown in the `Preview Any` node so you can fold the regions into the prompt.
+- **Boxes reach the model through `Flux3LayoutPrompt`.** `Create Bounding Boxes` puts the boxes on its `elements` output, and `Flux3LayoutPrompt` appends them to the prompt as a JSON layout table with an `id`, a `bbox`, and a `desc` per box. Boxes are numbered `obj_1`, `obj_2` and so on in the order you draw them, so draw the background element first. Cite a box in the prompt as `<obj_1>` to tie the words to that box. With no boxes drawn, the node returns the prompt unchanged.
+- **The image edit templates already use the boxes.** They send `Flux3LayoutPrompt` through the prompt upsampler and a switch into the sampler text encode, so the box list reaches the model without rewiring. Write the edit sentence in the layout node's own prompt field and keep `<obj_1>` where the box belongs.
+- **The text to image templates leave the layout output unconnected.** It goes to `Preview Any` only. Type the prompt in `Flux3LayoutPrompt`, then drag its `prompt` output onto the subgraph prompt input. Turn `enable_prompt_upsampler` off first, or the Qwen3.5 upsampler rewrites that text, layout table included, before sampling.
+- **For text to image, cover the whole canvas with one box.** Write the scene in that box's description. The model then gets a placement row for the entire frame instead of text with no layout information.
+- **Keep the box canvas at the output aspect ratio.** Box coordinates are stored as fractions of the `width` and `height` of `Create Bounding Boxes`, so a canvas with a different aspect ratio than the output latent stretches the regions. Any canvas size lines up as long as the ratio matches. The shipped image edit templates draw on a 896 x 1152 canvas while their latent is 1024 x 1024, so change the latent to match the box canvas when you need exact regions.
+- **Colors come from the box description.** The palette in the box editor is ignored, so name colors in the description.
diff --git a/zh/tutorials/flux/flux-3-klein.mdx b/zh/tutorials/flux/flux-3-klein.mdx
index 88d2029beb..f0a1cf4fe1 100644
--- a/zh/tutorials/flux/flux-3-klein.mdx
+++ b/zh/tutorials/flux/flux-3-klein.mdx
@@ -2,14 +2,14 @@
 title: "FLUX 3 Image Klein ComfyUI 指南"
 description: "在 ComfyUI 本地运行 FLUX.3 [klein]：开源权重文生图与指令驱动图像编辑，并提供 4–16 步的蒸馏版快速采样。"
 sidebarTitle: "FLUX 3 Klein"
-translationSourceHash: 7cedfbb4
+translationSourceHash: 1011d294
 translationFrom: tutorials/flux/flux-3-klein.mdx
 translationBlockHashes:
   "_intro": 3e9c5a21
   "About FLUX.3 [klein]": 1d7659ce
   "FLUX 3 Image Klein Workflows": d0941358
-  "FLUX 3 Image Klein Model Downloads": 0170574c
-  "Prompting and reference tips": 43ad4a36
+  "FLUX 3 Image Klein Model Downloads": 2972cb13
+  "Prompting and reference tips": 9f809dad
 ---
 import UpdateReminder from "/snippets/zh/tutorials/update-reminder.mdx"
 
@@ -175,7 +175,7 @@ FLUX 3 Klein 的权重存放在需要同意许可的仓库 [black-forest-labs/fl
     两个 checkpoint 共用的文本编码器。使用 `CLIPLoader` 类型 `flux3` 加载。
   </Card>
 <Card title="qwen3.5_9b_flux3_prompt_upsampler_w4a8.safetensors" icon="download" href="https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy">
-    文生图模板使用的提示词上采样器。
+    短提示词用的提示词上采样器。四个模板都会加载它。
   </Card>
 <Card title="flux2-vae.safetensors" icon="download" href="https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy">
     FLUX 3 Klein 的 VAE，即 FLUX.2 VAE。
@@ -203,4 +203,9 @@ FLUX 3 Klein 的权重存放在需要同意许可的仓库 [black-forest-labs/fl
 - **参考图会被加入两个条件中。** 按顺序把额外图像连接到 `Set Reference Multi Latent`，用 `<ref_image_0>`、`<ref_image_1>` 等来引用它们，并像参考 pipeline 那样把每一张同时加入正面和负面提示词。
 - **参考图尺寸是一档成本调节。** 模板会把每张参考图调整大小到约 100 万像素，输出 Latent 也采用该尺寸。保持参考图的原始尺寸（Black Forest Labs 自家的 pipeline 最高可到约 1600 万像素）能让编辑保留更多细节，但成本也更高。
 - **步数取决于 checkpoint。** 在模板中，完整 checkpoint 以 50 步运行，文生图引导为 3.5，图像编辑为 1。步数蒸馏版 checkpoint 以 16 步运行，引导固定为 1，想要快速跑一遍时就用它。
-- **模板自带 `Create Bounding Boxes` 节点。** 它会在来源图像上绘制边界框，用于规划区域编辑，其边界框列表显示在 `Preview Any` 节点中，方便你把这些区域融入提示词。
+- **边界框通过 `Flux3LayoutPrompt` 传给模型。** `Create Bounding Boxes` 把框输出到 `elements` 输出上，`Flux3LayoutPrompt` 再把它们以 JSON 布局表的形式追加到提示词后面，每个框包含 `id`、`bbox` 和 `desc`。框按绘制顺序编号为 `obj_1`、`obj_2`，所以先画背景元素。在提示词里用 `<obj_1>` 引用即可把文字绑定到该框。没有绘制任何框时，该节点原样返回提示词。
+- **图像编辑模板已经用上了边界框。** 它们把 `Flux3LayoutPrompt` 经过提示词上采样器和一个开关接到采样器的文本编码上，边界框列表无需重新连线就能进入模型。把编辑指令写在这个布局节点自己的提示词栏里，并在框对应的位置保留 `<obj_1>`。
+- **文生图模板的布局输出默认未连接。** 它只连到 `Preview Any`。在 `Flux3LayoutPrompt` 里写好提示词，再把它的 `prompt` 输出拖到子图的提示词输入上。先把 `enable_prompt_upsampler` 关掉，否则 Qwen3.5 上采样器会在采样前重写这段文字，布局表也一起被改写。
+- **文生图时用一个框覆盖整张画布。** 把场景描述写进这个框的描述里。这样模型拿到的是一行覆盖全画面的位置信息，而不是没有任何布局信息的文字。
+- **框的画布要与输出保持同一宽高比。** 框坐标按 `Create Bounding Boxes` 的 `width` 和 `height` 的比例存储，画布宽高比与输出 Latent 不一致时区域会被拉伸。只要比例一致，画布多大都能对齐。图像编辑模板的画布是 896 x 1152，而其 Latent 是 1024 x 1024；需要精确区域时把 Latent 改成与画布一致的比例。
+- **颜色写在框的描述里。** 编辑器中的调色板会被忽略，所以把颜色写进描述。

From e148373c8eeaf8739120e4d022db888b0284b4c6 Mon Sep 17 00:00:00 2001
From: lin-bot23 <linzhixiong23@gmail.com>
Date: Wed, 28 Oct 2026 18:19:19 +0800
Subject: [PATCH 05/13] docs(flux): correct the switch branch and the output
 size on the Klein page

---
 es/tutorials/flux/flux-3-klein.mdx | 12 ++++++------
 ja/tutorials/flux/flux-3-klein.mdx | 12 ++++++------
 ko/tutorials/flux/flux-3-klein.mdx | 12 ++++++------
 tutorials/flux/flux-3-klein.mdx    |  6 +++---
 zh/tutorials/flux/flux-3-klein.mdx | 12 ++++++------
 5 files changed, 27 insertions(+), 27 deletions(-)

diff --git a/es/tutorials/flux/flux-3-klein.mdx b/es/tutorials/flux/flux-3-klein.mdx
index 8135ca5890..c69f4305dd 100644
--- a/es/tutorials/flux/flux-3-klein.mdx
+++ b/es/tutorials/flux/flux-3-klein.mdx
@@ -2,14 +2,14 @@
 title: "FLUX 3 Image Klein en ComfyUI: texto a imagen y edición"
 description: "Ejecuta FLUX.3 [klein] en local en ComfyUI: texto a imagen con pesos abiertos y edición por instrucciones, con un checkpoint destilado para 4–16 pasos."
 sidebarTitle: "FLUX 3 Klein"
-translationSourceHash: 1011d294
+translationSourceHash: e78fedba
 translationFrom: tutorials/flux/flux-3-klein.mdx
 translationBlockHashes:
   "_intro": 3e9c5a21
   "About FLUX.3 [klein]": 1d7659ce
-  "FLUX 3 Image Klein Workflows": d0941358
+  "FLUX 3 Image Klein Workflows": 151dcb06
   "FLUX 3 Image Klein Model Downloads": 2972cb13
-  "Prompting and reference tips": 9f809dad
+  "Prompting and reference tips": 8bc95e3b
 ---
 import UpdateReminder from "/snippets/es/tutorials/update-reminder.mdx"
 
@@ -112,7 +112,7 @@ Sube este archivo al nodo `LoadImage` correspondiente:
 1. Carga tu imagen fuente en el nodo `LoadImage`
 2. Conéctala a `Set Reference Multi Latent`. La primera referencia es `<ref_image_0>` en el prompt, la siguiente es `<ref_image_1>`, y así sucesivamente
 3. Escribe la edición como una descripción completa del resultado. En el ejemplo incluido el prompt es `In <ref_image_0>, change the dress color <obj_1> to red`
-4. Ajusta el lienzo con `Empty Flux 2 Latent` (1024 x 1024 en la plantilla) si quieres un tamaño de salida diferente
+4. La salida toma la relación de aspecto de la referencia a 1 megapíxel. Sube `megapixels` en `Set Reference Multi Latent` para una imagen mayor
 5. Haz clic en **Queue**. `ImageCompare` muestra la entrada y el resultado uno al lado del otro
 
 **Salida de ejemplo**
@@ -204,8 +204,8 @@ Las plantillas cargan las compilaciones int8 para mantener pequeño el consumo d
 - **El tamaño de la referencia es un control de costo.** Las plantillas redimensionan cada referencia a aproximadamente 1 megapíxel y el latente de salida toma ese tamaño. Mantener las referencias en su tamaño nativo, que es lo que hace el pipeline propio de Black Forest Labs hasta unos 16 megapíxeles, conserva más detalle en la edición a un costo mayor.
 - **La cantidad de pasos depende del checkpoint.** El checkpoint completo se ejecuta con 50 pasos en las plantillas, con orientación 3.5 en texto a imagen y 1 en edición de imagen. El checkpoint destilado por pasos se ejecuta con 16 pasos y orientación fija en 1, y es el que debes usar cuando quieras una pasada rápida.
 - **Las cajas llegan al modelo a través de `Flux3LayoutPrompt`.** `Create Bounding Boxes` entrega las cajas en su salida `elements`, y `Flux3LayoutPrompt` las añade al prompt como una tabla de diseño JSON con un `id`, un `bbox` y un `desc` por caja. Las cajas se numeran `obj_1`, `obj_2` y así sucesivamente en el orden en que las dibujas, así que dibuja primero el elemento de fondo. Cita una caja en el prompt como `<obj_1>` para vincular las palabras a esa caja. Si no hay cajas dibujadas, el nodo devuelve el prompt sin cambios.
-- **Las plantillas de edición de imagen ya usan las cajas.** Llevan `Flux3LayoutPrompt` a través del upsampler de prompt y un switch hasta el codificador de texto del sampler, así que la lista de cajas llega al modelo sin recablear nada. Escribe la frase de edición en el campo de prompt del propio nodo de diseño y deja `<obj_1>` donde corresponde la caja.
+- **Las plantillas de edición de imagen llevan las cajas por un switch.** `Flux3LayoutPrompt` alimenta `ComfySwitchNode`, y su salida es el prompt de texto que recibe el sampler. El switch viene en la rama del upsampler, cuyo system prompt pide prosa plana sin JSON, coordenadas ni tokens entre corchetes angulares, así que la tabla de diseño se reescribe. Pon el switch en `true` para enviar el prompt de diseño al modelo tal como está escrito.
 - **Las plantillas de texto a imagen dejan la salida de diseño sin conectar.** Solo llega a `Preview Any`. Escribe el prompt en `Flux3LayoutPrompt` y arrastra su salida `prompt` hasta la entrada de prompt del subgrafo. Desactiva `enable_prompt_upsampler` antes: si no, el upsampler Qwen3.5 reescribe ese texto, tabla de diseño incluida, antes de muestrear.
 - **En texto a imagen, cubre todo el lienzo con una sola caja.** Escribe la escena en la descripción de esa caja. El modelo recibe entonces una fila de colocación para todo el encuadre en lugar de texto sin información de diseño.
-- **Mantén el lienzo de las cajas con la relación de aspecto de la salida.** Las coordenadas se guardan como fracciones del `width` y el `height` de `Create Bounding Boxes`, así que un lienzo con una relación de aspecto distinta a la del latente de salida estira las regiones. Cualquier tamaño de lienzo cuadra si la proporción coincide. Las plantillas de edición dibujan sobre un lienzo de 896 x 1152 y su latente es de 1024 x 1024, así que cambia el tamaño del latente al del lienzo cuando necesites regiones exactas.
+- **Mantén el lienzo de las cajas con la relación de aspecto de la salida.** Las coordenadas se guardan como fracciones del `width` y el `height` de `Create Bounding Boxes`, así que lo que debe coincidir es la proporción, no el tamaño en píxeles. En las plantillas de edición la salida sigue a la referencia: `Set Reference Multi Latent` redimensiona la primera referencia a 1 megapíxel y la devuelve como `empty_latent`, y el nodo `Empty Flux 2 Latent` de esas plantillas no está conectado. Ajusta el lienzo de las cajas a la referencia.
 - **Los colores van en la descripción de la caja.** La paleta del editor se ignora, así que nombra los colores en la descripción.
diff --git a/ja/tutorials/flux/flux-3-klein.mdx b/ja/tutorials/flux/flux-3-klein.mdx
index aa2d32b35c..c424b71482 100644
--- a/ja/tutorials/flux/flux-3-klein.mdx
+++ b/ja/tutorials/flux/flux-3-klein.mdx
@@ -2,14 +2,14 @@
 title: "FLUX 3 Image Klein ComfyUI ガイド"
 description: "ComfyUI で FLUX.3 [klein] をローカル実行: オープンウェイトのテキストから画像生成と指示による画像編集。4〜16 ステップの蒸留版にも対応。"
 sidebarTitle: "FLUX 3 Klein"
-translationSourceHash: 1011d294
+translationSourceHash: e78fedba
 translationFrom: tutorials/flux/flux-3-klein.mdx
 translationBlockHashes:
   "_intro": 3e9c5a21
   "About FLUX.3 [klein]": 1d7659ce
-  "FLUX 3 Image Klein Workflows": d0941358
+  "FLUX 3 Image Klein Workflows": 151dcb06
   "FLUX 3 Image Klein Model Downloads": 2972cb13
-  "Prompting and reference tips": 9f809dad
+  "Prompting and reference tips": 8bc95e3b
 ---
 import UpdateReminder from "/snippets/ja/tutorials/update-reminder.mdx"
 
@@ -112,7 +112,7 @@ FLUX.3 [klein] は、Black Forest Labs によるローカル ComfyUI 向けの
 1. `LoadImage` ノードでソース画像を読み込む
 2. それを `Set Reference Multi Latent` に接続します。最初の参照はプロンプト内の `<ref_image_0>`、次は `<ref_image_1>`、以降も同様です
 3. 編集内容を結果の完全な説明として記述します。付属の例のプロンプトは `In <ref_image_0>, change the dress color <obj_1> to red` です
-4. 別の出力サイズが必要な場合は、`Empty Flux 2 Latent`（テンプレートでは 1024 x 1024）でキャンバスを調整します
+4. 出力は参照画像のアスペクト比で約 1 メガピクセルになります。大きくしたい場合は `Set Reference Multi Latent` の `megapixels` を上げます
 5. **Queue** をクリックします。`ImageCompare` は入力と結果を並べて表示します
 
 **出力例**
@@ -204,8 +204,8 @@ FLUX 3 Klein の重みは、ライセンスへの同意が必要なリポジト
 - **参照のサイズはコストの調整ダイヤルです。** テンプレートは各参照を約1メガピクセルにリサイズし、出力潜在はそのサイズになります。Black Forest Labs 自身のパイプラインが最大約16メガピクセルまで対応しているように、参照をネイティブサイズのままにしておくと、より高いコストと引き換えに編集でより多くのディテールを保てます。
 - **ステップ数はチェックポイントによって異なります。** テンプレートでは、フルのチェックポイントは50ステップで実行され、テキストから画像ではガイダンス3.5、画像編集では1になります。ステップ蒸留版のチェックポイントは16ステップ、ガイダンスは1に固定で実行され、高速に処理したいときに使うのはこちらです。
 - **ボックスは `Flux3LayoutPrompt` 経由でモデルに渡ります。** `Create Bounding Boxes` はボックスを `elements` 出力に出し、`Flux3LayoutPrompt` がそれを `id`、`bbox`、`desc` を持つ JSON レイアウト表としてプロンプトの末尾に追加します。ボックスは描いた順に `obj_1`、`obj_2` と番号が振られるので、背景の要素を先に描いてください。プロンプト内で `<obj_1>` として参照すると、その語がそのボックスに紐づきます。ボックスを描いていない場合、このノードはプロンプトをそのまま返します。
-- **画像編集テンプレートはすでにボックスを使っています。** `Flux3LayoutPrompt` をプロンプトアップサンプラとスイッチ経由でサンプラーのテキストエンコードに接続しているため、配線を変えなくてもボックス一覧がモデルに届きます。編集の文はこのレイアウトノード自身のプロンプト欄に書き、ボックスに対応する位置に `<obj_1>` を残します。
+- **画像編集テンプレートはスイッチ経由でボックスを渡します。** `Flux3LayoutPrompt` は `ComfySwitchNode` に接続され、その出力がサンプラーに渡るテキストプロンプトになります。スイッチはアップサンプラ側が既定で、そのシステムプロンプトは JSON、座標、山かっこ付きのトークンを含めないプレーンな文章を求めるため、レイアウト表は書き換えられます。`true` に切り替えると、レイアウトプロンプトがそのままモデルに渡ります。
 - **テキストから画像のテンプレートはレイアウト出力が未接続です。** `Preview Any` にしかつながっていません。`Flux3LayoutPrompt` にプロンプトを書き、その `prompt` 出力をサブグラフのプロンプト入力にドラッグします。先に `enable_prompt_upsampler` を切ってください。切らないと Qwen3.5 のアップサンプラがサンプリング前にこのテキストを書き換え、レイアウト表も一緒に変わります。
 - **テキストから画像では、キャンバス全体を覆う 1 つのボックスを使います。** シーンの説明はそのボックスの説明欄に書きます。モデルはレイアウト情報のないテキストではなく、フレーム全体に対する配置行を受け取ります。
-- **ボックスのキャンバスは出力と同じアスペクト比にします。** ボックス座標は `Create Bounding Boxes` の `width` と `height` に対する比率で保存されるため、キャンバスのアスペクト比が出力潜在と異なると領域が引き伸ばされます。比率が同じならキャンバスのピクセル数はいくつでも合います。画像編集テンプレートのキャンバスは 896 x 1152、潜在は 1024 x 1024 なので、正確な領域が必要なときは潜在をキャンバスと同じ比率に変更してください。
+- **ボックスのキャンバスは出力と同じアスペクト比にします。** ボックス座標は `Create Bounding Boxes` の `width` と `height` に対する比率で保存されるため、合わせるのは比率で、ピクセル数ではありません。画像編集テンプレートの出力は参照画像に従います。`Set Reference Multi Latent` が最初の参照を 1 メガピクセルにリサイズして `empty_latent` として返し、そこにある `Empty Flux 2 Latent` ノードは接続されていません。ボックスのキャンバスを参照画像に合わせてください。
 - **色はボックスの説明欄に書きます。** エディタのパレットは無視されるため、色は説明に書いてください。
diff --git a/ko/tutorials/flux/flux-3-klein.mdx b/ko/tutorials/flux/flux-3-klein.mdx
index 45f7105162..b5eb8228c2 100644
--- a/ko/tutorials/flux/flux-3-klein.mdx
+++ b/ko/tutorials/flux/flux-3-klein.mdx
@@ -2,14 +2,14 @@
 title: "FLUX 3 Image Klein ComfyUI 가이드"
 description: "ComfyUI에서 FLUX.3 [klein]을 로컬로 실행: 오픈 웨이트 텍스트 기반 이미지 생성과 지시 기반 이미지 편집, 4~16스텝 증류판 포함."
 sidebarTitle: "FLUX 3 Klein"
-translationSourceHash: 1011d294
+translationSourceHash: e78fedba
 translationFrom: tutorials/flux/flux-3-klein.mdx
 translationBlockHashes:
   "_intro": 3e9c5a21
   "About FLUX.3 [klein]": 1d7659ce
-  "FLUX 3 Image Klein Workflows": d0941358
+  "FLUX 3 Image Klein Workflows": 151dcb06
   "FLUX 3 Image Klein Model Downloads": 2972cb13
-  "Prompting and reference tips": 9f809dad
+  "Prompting and reference tips": 8bc95e3b
 ---
 import UpdateReminder from "/snippets/ko/tutorials/update-reminder.mdx"
 
@@ -112,7 +112,7 @@ FLUX.3 [klein]은 Black Forest Labs의 로컬 ComfyUI용 오픈 웨이트 FLUX.3
 1. `LoadImage` 노드에서 소스 이미지를 불러옵니다
 2. `Set Reference Multi Latent`에 연결합니다. 첫 번째 참조는 프롬프트에서 `<ref_image_0>`이고, 다음은 `<ref_image_1>`이며, 이런 식으로 계속 이어집니다
 3. 편집 내용을 결과에 대한 완전한 설명으로 작성합니다. 제공된 예시의 프롬프트는 `In <ref_image_0>, change the dress color <obj_1> to red`입니다
-4. 다른 출력 크기를 원한다면 `Empty Flux 2 Latent`(템플릿에서는 1024 x 1024)로 캔버스를 조정합니다
+4. 출력은 레퍼런스 이미지의 가로세로 비율로 약 1메가픽셀입니다. 더 크게 만들려면 `Set Reference Multi Latent`의 `megapixels`를 올리세요
 5. **실행 대기열**을 클릭합니다. `ImageCompare`가 입력과 결과를 나란히 보여줍니다
 
 **예시 출력**
@@ -204,8 +204,8 @@ FLUX 3 Klein 가중치는 라이선스 동의가 필요한 저장소 [black-fore
 - **레퍼런스 크기는 비용을 조절하는 다이얼입니다.** 템플릿은 각 레퍼런스의 크기를 약 1메가픽셀로 조정하며, 출력 잠재 데이터는 그 크기를 갖습니다. Black Forest Labs 자체 pipeline이 약 16메가픽셀까지 유지하는 것처럼 레퍼런스를 원본 크기로 유지하면 비용이 더 높아지는 대신 편집 결과에 더 많은 디테일이 남습니다.
 - **스텝 수는 checkpoint에 따라 다릅니다.** 전체 checkpoint는 템플릿에서 50 스텝으로 실행되며, 텍스트 기반 생성에서는 가이드 3.5, 이미지 편집에서는 1입니다. 스텝 증류(step-distilled) checkpoint는 16 스텝에 가이드가 1로 고정되어 실행되며, 빠르게 처리하고 싶을 때 사용하는 것이 바로 이 checkpoint입니다.
 - **박스는 `Flux3LayoutPrompt`를 거쳐 모델에 전달됩니다.** `Create Bounding Boxes`는 박스를 `elements` 출력으로 내보내고, `Flux3LayoutPrompt`는 이를 `id`, `bbox`, `desc`를 담은 JSON 레이아웃 표로 프롬프트 뒤에 덧붙입니다. 박스는 그린 순서대로 `obj_1`, `obj_2`로 번호가 붙으므로 배경 요소를 먼저 그리세요. 프롬프트에서 `<obj_1>`로 인용하면 해당 단어가 그 박스에 연결됩니다. 박스를 그리지 않으면 이 노드는 프롬프트를 그대로 반환합니다.
-- **이미지 편집 템플릿은 이미 박스를 사용합니다.** `Flux3LayoutPrompt`를 프롬프트 업샘플러와 스위치를 거쳐 샘플러의 텍스트 인코드로 보내므로, 배선을 바꾸지 않아도 박스 목록이 모델에 전달됩니다. 편집 문장은 이 레이아웃 노드 자체의 프롬프트 칸에 쓰고, 박스에 해당하는 자리에 `<obj_1>`을 남겨 두세요.
+- **이미지 편집 템플릿은 스위치를 거쳐 박스를 전달합니다.** `Flux3LayoutPrompt`는 `ComfySwitchNode`에 연결되고, 그 출력이 샘플러에 들어가는 텍스트 프롬프트가 됩니다. 스위치는 업샘플러 쪽이 기본이며, 그 시스템 프롬프트는 JSON, 좌표, 꺾쇠 토큰이 없는 평문을 요구하므로 레이아웃 표가 다시 쓰입니다. 스위치를 `true`로 바꾸면 레이아웃 프롬프트가 그대로 모델에 전달됩니다.
 - **텍스트에서 이미지 템플릿은 레이아웃 출력이 연결되어 있지 않습니다.** `Preview Any`에만 연결됩니다. `Flux3LayoutPrompt`에 프롬프트를 쓰고 그 `prompt` 출력을 서브그래프의 프롬프트 입력으로 끌어다 놓으세요. 먼저 `enable_prompt_upsampler`를 끄세요. 끄지 않으면 Qwen3.5 업샘플러가 샘플링 전에 이 텍스트를 다시 쓰면서 레이아웃 표도 함께 바뀝니다.
 - **텍스트에서 이미지 생성 시 캔버스 전체를 덮는 박스 하나를 사용하세요.** 장면 설명은 그 박스의 설명에 적습니다. 그러면 모델은 레이아웃 정보가 없는 텍스트 대신 프레임 전체에 대한 배치 행을 받습니다.
-- **박스 캔버스는 출력과 같은 가로세로 비율로 맞추세요.** 박스 좌표는 `Create Bounding Boxes`의 `width`와 `height`에 대한 비율로 저장되므로, 캔버스 비율이 출력 잠재 데이터와 다르면 영역이 늘어납니다. 비율만 같으면 캔버스 픽셀 크기는 상관없습니다. 이미지 편집 템플릿의 캔버스는 896 x 1152이고 잠재 데이터는 1024 x 1024이므로, 정확한 영역이 필요하면 잠재 크기를 캔버스 비율에 맞게 바꾸세요.
+- **박스 캔버스는 출력과 같은 가로세로 비율로 맞추세요.** 박스 좌표는 `Create Bounding Boxes`의 `width`와 `height`에 대한 비율로 저장되므로, 맞춰야 하는 것은 비율이지 픽셀 크기가 아닙니다. 이미지 편집 템플릿의 출력은 레퍼런스 이미지를 따릅니다. `Set Reference Multi Latent`가 첫 레퍼런스를 1메가픽셀로 조정해 `empty_latent`로 반환하며, 그 템플릿의 `Empty Flux 2 Latent` 노드는 연결되어 있지 않습니다. 박스 캔버스를 레퍼런스에 맞추세요.
 - **색상은 박스 설명에 적습니다.** 편집기의 팔레트는 무시되므로 색을 설명에 적으세요.
diff --git a/tutorials/flux/flux-3-klein.mdx b/tutorials/flux/flux-3-klein.mdx
index d080b6fbbf..d5dd1140e8 100644
--- a/tutorials/flux/flux-3-klein.mdx
+++ b/tutorials/flux/flux-3-klein.mdx
@@ -105,7 +105,7 @@ Upload this file to the matching `LoadImage` node:
 1. Load your source image in the `LoadImage` node
 2. Connect it to `Set Reference Multi Latent`. The first reference is `<ref_image_0>` in the prompt, the next is `<ref_image_1>`, and so on
 3. Write the edit as a full description of the result. In the shipped example the prompt is `In <ref_image_0>, change the dress color <obj_1> to red`
-4. Adjust the canvas with `Empty Flux 2 Latent` (1024 x 1024 in the template) if you want a different output size
+4. The output takes the reference's aspect ratio at 1 megapixel. Raise `megapixels` in `Set Reference Multi Latent` for a larger image
 5. Click **Queue**. `ImageCompare` shows the input and the result side by side
 
 **Example output**
@@ -197,8 +197,8 @@ The templates load the int8 builds to keep the memory footprint small. With more
 - **Reference size is a cost dial.** The templates resize each reference to roughly 1 megapixel and the output latent takes that size. Keeping references at their native size, which Black Forest Labs' own pipeline does up to about 16 megapixels, keeps more detail in the edit at a higher cost.
 - **Step count depends on the checkpoint.** The full checkpoint runs at 50 steps in the templates, with guidance 3.5 on text to image and 1 on image edit. The step-distilled checkpoint runs at 16 steps with guidance fixed at 1, and it is the one to use when you want a fast pass.
 - **Boxes reach the model through `Flux3LayoutPrompt`.** `Create Bounding Boxes` puts the boxes on its `elements` output, and `Flux3LayoutPrompt` appends them to the prompt as a JSON layout table with an `id`, a `bbox`, and a `desc` per box. Boxes are numbered `obj_1`, `obj_2` and so on in the order you draw them, so draw the background element first. Cite a box in the prompt as `<obj_1>` to tie the words to that box. With no boxes drawn, the node returns the prompt unchanged.
-- **The image edit templates already use the boxes.** They send `Flux3LayoutPrompt` through the prompt upsampler and a switch into the sampler text encode, so the box list reaches the model without rewiring. Write the edit sentence in the layout node's own prompt field and keep `<obj_1>` where the box belongs.
+- **The image edit templates route the boxes through a switch.** `Flux3LayoutPrompt` feeds `ComfySwitchNode`, and the switch output is the text prompt at the sampler. The switch ships on the upsampler branch, whose system prompt asks for plain prose with no JSON, coordinates, or angle-bracket tokens, so the box table gets rewritten. Set the switch to `true` to send the layout prompt to the model as written.
 - **The text to image templates leave the layout output unconnected.** It goes to `Preview Any` only. Type the prompt in `Flux3LayoutPrompt`, then drag its `prompt` output onto the subgraph prompt input. Turn `enable_prompt_upsampler` off first, or the Qwen3.5 upsampler rewrites that text, layout table included, before sampling.
 - **For text to image, cover the whole canvas with one box.** Write the scene in that box's description. The model then gets a placement row for the entire frame instead of text with no layout information.
-- **Keep the box canvas at the output aspect ratio.** Box coordinates are stored as fractions of the `width` and `height` of `Create Bounding Boxes`, so a canvas with a different aspect ratio than the output latent stretches the regions. Any canvas size lines up as long as the ratio matches. The shipped image edit templates draw on a 896 x 1152 canvas while their latent is 1024 x 1024, so change the latent to match the box canvas when you need exact regions.
+- **Keep the box canvas at the output aspect ratio.** Box coordinates are stored as fractions of the `width` and `height` of `Create Bounding Boxes`, so the ratio is what has to match, not the pixel size. In the image edit templates the output follows the reference: `Set Reference Multi Latent` resizes the first reference to 1 megapixel and returns it as `empty_latent`, and the `Empty Flux 2 Latent` node there is not connected. Match the box canvas to the reference.
 - **Colors come from the box description.** The palette in the box editor is ignored, so name colors in the description.
diff --git a/zh/tutorials/flux/flux-3-klein.mdx b/zh/tutorials/flux/flux-3-klein.mdx
index f0a1cf4fe1..3eecd174a4 100644
--- a/zh/tutorials/flux/flux-3-klein.mdx
+++ b/zh/tutorials/flux/flux-3-klein.mdx
@@ -2,14 +2,14 @@
 title: "FLUX 3 Image Klein ComfyUI 指南"
 description: "在 ComfyUI 本地运行 FLUX.3 [klein]：开源权重文生图与指令驱动图像编辑，并提供 4–16 步的蒸馏版快速采样。"
 sidebarTitle: "FLUX 3 Klein"
-translationSourceHash: 1011d294
+translationSourceHash: e78fedba
 translationFrom: tutorials/flux/flux-3-klein.mdx
 translationBlockHashes:
   "_intro": 3e9c5a21
   "About FLUX.3 [klein]": 1d7659ce
-  "FLUX 3 Image Klein Workflows": d0941358
+  "FLUX 3 Image Klein Workflows": 151dcb06
   "FLUX 3 Image Klein Model Downloads": 2972cb13
-  "Prompting and reference tips": 9f809dad
+  "Prompting and reference tips": 8bc95e3b
 ---
 import UpdateReminder from "/snippets/zh/tutorials/update-reminder.mdx"
 
@@ -112,7 +112,7 @@ FLUX.3 [klein] 是 Black Forest Labs 面向本地 ComfyUI 的开源权重 FLUX.3
 1. 在 `LoadImage` 节点中加载来源图像
 2. 将其连接到 `Set Reference Multi Latent`。第一个参考在提示词中为 `<ref_image_0>`，下一个为 `<ref_image_1>`，依此类推
 3. 将编辑内容写为对结果的完整描述。随附的示例中，提示词为 `In <ref_image_0>, change the dress color <obj_1> to red`
-4. 如果想要不同的输出尺寸，可使用 `Empty Flux 2 Latent` 调整画布（模板中为 1024 x 1024）
+4. 输出尺寸为参考图的比例，约 100 万像素。想要更大的图像就调高 `Set Reference Multi Latent` 里的 `megapixels`
 5. 点击 **队列**。`ImageCompare` 会并排显示输入和结果
 
 **示例输出**
@@ -204,8 +204,8 @@ FLUX 3 Klein 的权重存放在需要同意许可的仓库 [black-forest-labs/fl
 - **参考图尺寸是一档成本调节。** 模板会把每张参考图调整大小到约 100 万像素，输出 Latent 也采用该尺寸。保持参考图的原始尺寸（Black Forest Labs 自家的 pipeline 最高可到约 1600 万像素）能让编辑保留更多细节，但成本也更高。
 - **步数取决于 checkpoint。** 在模板中，完整 checkpoint 以 50 步运行，文生图引导为 3.5，图像编辑为 1。步数蒸馏版 checkpoint 以 16 步运行，引导固定为 1，想要快速跑一遍时就用它。
 - **边界框通过 `Flux3LayoutPrompt` 传给模型。** `Create Bounding Boxes` 把框输出到 `elements` 输出上，`Flux3LayoutPrompt` 再把它们以 JSON 布局表的形式追加到提示词后面，每个框包含 `id`、`bbox` 和 `desc`。框按绘制顺序编号为 `obj_1`、`obj_2`，所以先画背景元素。在提示词里用 `<obj_1>` 引用即可把文字绑定到该框。没有绘制任何框时，该节点原样返回提示词。
-- **图像编辑模板已经用上了边界框。** 它们把 `Flux3LayoutPrompt` 经过提示词上采样器和一个开关接到采样器的文本编码上，边界框列表无需重新连线就能进入模型。把编辑指令写在这个布局节点自己的提示词栏里，并在框对应的位置保留 `<obj_1>`。
+- **图像编辑模板通过一个开关接入边界框。** `Flux3LayoutPrompt` 接到 `ComfySwitchNode`，开关输出就是采样器使用的文本提示词。开关默认选中上采样分支，而该分支的系统提示词要求输出不含 JSON、坐标和尖括号标记的纯文本，所以布局表会被改写。把开关设为 `true`，布局提示词就会原样送入模型。
 - **文生图模板的布局输出默认未连接。** 它只连到 `Preview Any`。在 `Flux3LayoutPrompt` 里写好提示词，再把它的 `prompt` 输出拖到子图的提示词输入上。先把 `enable_prompt_upsampler` 关掉，否则 Qwen3.5 上采样器会在采样前重写这段文字，布局表也一起被改写。
 - **文生图时用一个框覆盖整张画布。** 把场景描述写进这个框的描述里。这样模型拿到的是一行覆盖全画面的位置信息，而不是没有任何布局信息的文字。
-- **框的画布要与输出保持同一宽高比。** 框坐标按 `Create Bounding Boxes` 的 `width` 和 `height` 的比例存储，画布宽高比与输出 Latent 不一致时区域会被拉伸。只要比例一致，画布多大都能对齐。图像编辑模板的画布是 896 x 1152，而其 Latent 是 1024 x 1024；需要精确区域时把 Latent 改成与画布一致的比例。
+- **框的画布要与输出保持同一宽高比。** 框坐标按 `Create Bounding Boxes` 的 `width` 和 `height` 的比例存储，所以要对齐的是比例，不是像素尺寸。图像编辑模板的输出跟随参考图：`Set Reference Multi Latent` 把第一张参考图缩放到 100 万像素并作为 `empty_latent` 输出，其中的 `Empty Flux 2 Latent` 节点并未连接。把框的画布对齐参考图即可。
 - **颜色写在框的描述里。** 编辑器中的调色板会被忽略，所以把颜色写进描述。

From 217efa36766e1ece29bae94fd98c38b276db9fa5 Mon Sep 17 00:00:00 2001
From: lin-bot23 <linzhixiong23@gmail.com>
Date: Thu, 29 Oct 2026 18:11:44 +0800
Subject: [PATCH 06/13] docs(minimax, ltx): move the H3 audio-first pass onto
 the native page and document the LTX-2.5 conv VAE

---
 es/tutorials/video/ltx/ltx-2-5.mdx            |  8 ++++--
 .../video/minimax/minimax-h3-native.mdx       | 26 +++++++++++++------
 ja/tutorials/video/ltx/ltx-2-5.mdx            |  8 ++++--
 .../video/minimax/minimax-h3-native.mdx       | 26 +++++++++++++------
 ko/tutorials/video/ltx/ltx-2-5.mdx            |  8 ++++--
 .../video/minimax/minimax-h3-native.mdx       | 26 +++++++++++++------
 tutorials/video/ltx/ltx-2-5.mdx               |  4 +++
 tutorials/video/minimax/minimax-h3-native.mdx | 16 +++++++++---
 zh/tutorials/video/ltx/ltx-2-5.mdx            |  8 ++++--
 .../video/minimax/minimax-h3-native.mdx       | 26 +++++++++++++------
 10 files changed, 113 insertions(+), 43 deletions(-)

diff --git a/es/tutorials/video/ltx/ltx-2-5.mdx b/es/tutorials/video/ltx/ltx-2-5.mdx
index 86b03673de..cc2b70256e 100644
--- a/es/tutorials/video/ltx/ltx-2-5.mdx
+++ b/es/tutorials/video/ltx/ltx-2-5.mdx
@@ -2,12 +2,12 @@
 title: "LTX-2.5 en ComfyUI: flujos de video, audio y multishot"
 description: "El modelo de video open source de Lightricks en ComfyUI: texto a video, imagen a video, primer-último fotograma, DFR, multishot, audio sincronizado y 4K HDR."
 sidebarTitle: "LTX-2.5"
-translationSourceHash: 4a7cea51
+translationSourceHash: ebef3f6f
 translationFrom: tutorials/video/ltx/ltx-2-5.mdx
 translationBlockHashes:
   "_intro": 4aa7d7de
   "Key features": 5bfa64be
-  "Getting started": b7968eb6
+  "Getting started": 7909413a
   "ComfyUI Native Workflows": a65c6f64
   "LTX-2.5 DFR: Image to Video": 3899f68b
   "Resources": f89b58b6
@@ -49,6 +49,10 @@ LTX-2.5 es compatible de forma nativa con ComfyUI. Para empezar:
   Los archivos del modelo LTX-2.5 están alojados en los repositorios restringidos [Lightricks/LTX-2.5](https://huggingface.co/Lightricks/LTX-2.5) y [Lightricks/LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler](https://huggingface.co/Lightricks/LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler) en Hugging Face. Abre las páginas de los repositorios, acepta la licencia del modelo y espera a que se apruebe tu solicitud de acceso antes de descargar los modelos. Las descargas de modelos fallarán si no tienes acceso.
 </Callout>
 
+<Note>
+LTX-2.5 incluye dos VAE de video que comparten el mismo espacio latente. Las plantillas cargan [ltx-2.5-video-vae-bf16.safetensors](https://huggingface.co/Lightricks/LTX-2.5/blob/main/vae/ltx-2.5-video-vae-bf16.safetensors), el decodificador de difusión. Aplica un paso de decodificación aprendido y no se puede descargar a CPU, así que una decodificación a tamaño completo ocupa mucha VRAM. Cuando esa decodificación se queda sin memoria, cambia `VAELoader` a [ltx-2.5-video-vae-conv-bf16.safetensors](https://huggingface.co/Lightricks/LTX-2.5/blob/main/vae/ltx-2.5-video-vae-conv-bf16.safetensors). Este archivo decodifica con el autocodificador convolucional, más rápido y ligero, con caras y texto en pantalla más suaves.
+</Note>
+
 ## Flujos de trabajo nativos de ComfyUI
 
 ### LTX-2.5 Texto a video (T2V)
diff --git a/es/tutorials/video/minimax/minimax-h3-native.mdx b/es/tutorials/video/minimax/minimax-h3-native.mdx
index dcc0f6cb27..5f1c134b4a 100644
--- a/es/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/es/tutorials/video/minimax/minimax-h3-native.mdx
@@ -2,16 +2,16 @@
 title: "Flujos de trabajo de MiniMax H3 en ComfyUI: texto a video, imagen a video y referencia a video"
 description: "Ejemplos paso a paso de flujos nativos de MiniMax H3 en ComfyUI, con anclaje de guías y máscaras de ruido latente para un control más fino del video."
 sidebarTitle: "Flujos nativos"
-translationSourceHash: 9e8d385c
+translationSourceHash: f8176106
 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx
 translationBlockHashes:
-  "_intro": 4f5274f5
-  "MiniMax H3 Text to Video (T2V)": 8807c872
-  "MiniMax H3 Image to Video (I2V)": c59554b0
+  "_intro": dde651e5
+  "MiniMax H3 Text to Video (T2V)": 42ec8cfb
+  "MiniMax H3 Image to Video (I2V)": 2a06e3fa
   "MiniMax H3 Reference to Video (R2V)": 961829df
-  "Advanced workflows with the native nodes": 37164e51
+  "Advanced workflows with the native nodes": 85460bff
 ---
-Esta página cubre las tres plantillas base de MiniMax H3 y las técnicas avanzadas con nodos nativos. Para conocer las capacidades, el ajuste y las mejoras de velocidad, consulta la [descripción general de MiniMax H3](/es/tutorials/video/minimax/minimax-h3).
+Esta página cubre las tres plantillas base de MiniMax H3 y las técnicas avanzadas con nodos nativos. Para conocer las capacidades y el licenciamiento, consulta la [descripción general de MiniMax H3](/es/tutorials/video/minimax/minimax-h3).
 
 ## MiniMax H3 Texto a Video (T2V)
 
@@ -73,7 +73,7 @@ ComfyUI/
 3. **Resolución**: el 16:9 nativo es **1344×768** (768 px en el lado corto); redondea a múltiplos de **32**
 4. **Duración**: se ajusta a **17k+5** fotogramas a 24 fps
 5. **Primer y último fotograma**: conecta las imágenes a `first_frame` / `last_frame` en `MiniMaxH3ImageToVideo` para FL2VA
-6. **Turbo**: **20** pasos por defecto; activa `turbo_mode` para el Lightning LoRA de **8** pasos (más rápido, con movimiento y audio algo más suaves). Sube los pasos (por ejemplo, **25**) para mejorar la calidad del movimiento
+6. **Turbo**: **20** pasos por defecto; activa `turbo_mode` para el Lightning LoRA de **8** pasos (más rápido, con movimiento y audio algo más suaves). Sube los pasos (por ejemplo, **25**) para mejorar la calidad del movimiento. Las acciones rápidas y los movimientos de cámara son lo primero que suaviza una programación turbo corta. Compara recuentos de pasos en un clip con movimiento real, no en un plano estático
 
 Guía base oficial: [guía de prompts](/es/tutorials/video/minimax/minimax-h3-prompt-guide).
 
@@ -144,7 +144,7 @@ ComfyUI/
 1. **Fotogramas clave**: `first_frame` / `last_frame` opcionales; el movimiento se genera entre ellos
 2. **Prompt**: planos, movimiento y audio en un solo bloque
 3. **Resolución / duración**: las mismas reglas que en T2V (objetivo nativo **1344×768**, cuadrícula **17k+5**)
-4. **Turbo**: `turbo_mode` usa `turbo_steps` (valor predeterminado **8**) y `turbo_model_strength` (valor predeterminado **1.0**) en lugar de 20 pasos
+4. **Turbo**: `turbo_mode` usa `turbo_steps` (valor predeterminado **8**) y `turbo_model_strength` (valor predeterminado **1.0**) en lugar de 20 pasos. Con una programación turbo corta, las acciones rápidas y los movimientos de cámara se suavizan primero, así que revísalos antes de quedarte con turbo
 
 `first_frame` se redimensiona al lienzo sin conservar su relación de aspecto, así que una imagen con otra relación de aspecto sale estirada. `last_frame`, en cambio, se recorta desde el centro. La plantilla de I2V se abre en **1:1** y **0.4** MP (**640×640**), así que ajusta la relación de aspecto a tu imagen y sube los megapíxeles para el render final.
 
@@ -257,3 +257,13 @@ Consulta la plantilla [Multiframe Reference](/es/tutorials/video/minimax/minimax
 Las [máscaras de ruido por token](https://github.com/Comfy-Org/ComfyUI/pull/15375) en los latentes de video y audio te permiten regenerar parte de un clip.
 
 Conecta una máscara a `denoise_mask` del sampler: **0** = mantener, **1** = regenerar. Las máscaras de video se ajustan a la cuadrícula de parches latentes de 2×2; las de audio, a frames latentes completos. Úsalas para relleno inteligente, eliminación o extensión mientras mantienes el resto estable.
+
+### Generar primero la banda sonora
+
+El audio y el video se desruidizan juntos, así que una programación corta o un LoRA turbo suavizan la pista junto con la imagen. Dos pasadas los separan: fija el audio en un lienzo pequeño y luego condiciona el video con él.
+
+1. **Pasada de audio barata**: Ejecuta una pasada de texto a video a **32×32** con `Boolean (Enable Lightning LoRA)` desactivado, **40–50** pasos y la duración que necesita el plano final. Los frames latentes de audio siguen la longitud del clip, no el tamaño del lienzo, así que esta pasada genera una pista de longitud completa por mucho menos que un render a resolución completa. Desconecta el Resolution Selector y escribe el tamaño en sus destinos `width` y `height`; el selector llega como mínimo a **0.1** megapíxeles. Los pasos están en el `PrimitiveInt` ajustado a **20** dentro del subgrafo (el que está en **6** es el recuento de pasos turbo). De esta pasada solo se conserva el audio.
+2. **Guardar la pista**: El subgrafo decodifica el audio dentro y expone solo el video, así que añade un nodo `SaveAudioAdvanced` dentro del subgrafo y conecta `VAEDecodeAudio` a él.
+3. **Condicionar la segunda pasada**: En el flujo R2V añade `LoadAudio`, conéctalo a `ref_audio` y mantén `audio_vae` conectado. Cita la pista como `<Audio 1>` e indica la relación: `reference` conserva timbre, ritmo y dicción sin copiar la señal, `fully_copy` la convierte en la pista final completa, `partially_copy` conserva parte de la línea de tiempo o capas seleccionadas, y `weak_reference` solo mantiene una similitud amplia. La segunda pasada parte así de una voz y un ritmo ya resueltos en lugar de inventar el habla desde cero. Para usar el mismo archivo, conecta `LoadAudio` a `CreateVideo` en vez de confiar en el prompt.
+
+El tamaño del lienzo intercambia calidad de voz por tiempo de render. Una pasada a 32×32 es casi gratis, pero a ese tamaño la voz puede sonar metálica. Sube el lienzo cuando la voz tenga que sostenerse en el corte final, y cuenta con que la pasada cueste más. El habla y las letras suelen sobrevivir al lienzo pequeño. Los efectos de sonido y la música atados a la acción en pantalla no, así que genéralos en la segunda pasada y combínalos con `partially_copy`.
diff --git a/ja/tutorials/video/ltx/ltx-2-5.mdx b/ja/tutorials/video/ltx/ltx-2-5.mdx
index b254f4009c..f3ba75b632 100644
--- a/ja/tutorials/video/ltx/ltx-2-5.mdx
+++ b/ja/tutorials/video/ltx/ltx-2-5.mdx
@@ -2,12 +2,12 @@
 title: "LTX-2.5: ComfyUI ビデオ・オーディオ・マルチショットワークフロー"
 description: "Lightricks のオープンウェイト動画モデルを ComfyUI で利用。テキストから動画、画像から動画、先頭・最終フレーム、DFR のワークフローに対応し、マルチショット、音声同期、4K HDR を備えます。"
 sidebarTitle: "LTX-2.5"
-translationSourceHash: 4a7cea51
+translationSourceHash: ebef3f6f
 translationFrom: tutorials/video/ltx/ltx-2-5.mdx
 translationBlockHashes:
   "_intro": 4aa7d7de
   "Key features": 5bfa64be
-  "Getting started": b7968eb6
+  "Getting started": 7909413a
   "ComfyUI Native Workflows": a65c6f64
   "LTX-2.5 DFR: Image to Video": 3899f68b
   "Resources": f89b58b6
@@ -49,6 +49,10 @@ LTX-2.5はComfyUIでネイティブにサポートされています。はじめ
   LTX-2.5 のモデルファイルは、Hugging Face のアクセス制限付き（gated）リポジトリ [Lightricks/LTX-2.5](https://huggingface.co/Lightricks/LTX-2.5) と [Lightricks/LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler](https://huggingface.co/Lightricks/LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler) でホストされています。リポジトリのページを開き、モデルライセンスに同意し、アクセス申請が承認されるまで待ってからモデルをダウンロードしてください。アクセスがないとモデルのダウンロードは失敗します。
 </Callout>
 
+<Note>
+LTX-2.5 には同じ latent 空間を共有する 2 種類の動画 VAE があります。テンプレートは拡散デコーダの [ltx-2.5-video-vae-bf16.safetensors](https://huggingface.co/Lightricks/LTX-2.5/blob/main/vae/ltx-2.5-video-vae-bf16.safetensors) を読み込みます。学習されたデコード段を通り、オフロードできないため、フルサイズのデコードは多くの VRAM を占有します。そのデコードでメモリが足りないときは、`VAELoader` を [ltx-2.5-video-vae-conv-bf16.safetensors](https://huggingface.co/Lightricks/LTX-2.5/blob/main/vae/ltx-2.5-video-vae-conv-bf16.safetensors) に切り替えてください。こちらは畳み込みオートエンコーダでデコードするため、より速く軽く、顔や画面内の文字はやや甘くなります。
+</Note>
+
 ## ComfyUI ネイティブワークフロー
 
 ### LTX-2.5 テキストからビデオ生成（T2V）
diff --git a/ja/tutorials/video/minimax/minimax-h3-native.mdx b/ja/tutorials/video/minimax/minimax-h3-native.mdx
index 3c775ec48a..8feae41a0b 100644
--- a/ja/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/ja/tutorials/video/minimax/minimax-h3-native.mdx
@@ -2,17 +2,17 @@
 title: "ComfyUI MiniMax H3 テキストから動画・画像から動画・参照から動画のワークフロー"
 description: "ComfyUI における MiniMax H3 ネイティブワークフローの手順例。ガイドのアンカー固定や latent ノイズマスクで生成動画をより細かく制御できます。"
 sidebarTitle: "ネイティブワークフロー"
-translationSourceHash: 9e8d385c
+translationSourceHash: f8176106
 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx
 translationBlockHashes:
-  "_intro": 4f5274f5
-  "MiniMax H3 Text to Video (T2V)": 8807c872
-  "MiniMax H3 Image to Video (I2V)": c59554b0
+  "_intro": dde651e5
+  "MiniMax H3 Text to Video (T2V)": 42ec8cfb
+  "MiniMax H3 Image to Video (I2V)": 2a06e3fa
   "MiniMax H3 Reference to Video (R2V)": 961829df
-  "Advanced workflows with the native nodes": 37164e51
+  "Advanced workflows with the native nodes": 85460bff
 ---
 
-このページは 3 つの基本 MiniMax H3 テンプレートとネイティブノードの高度なテクニックを扱います。機能、チューニング、高速化は [MiniMax H3 概要](/ja/tutorials/video/minimax/minimax-h3) を参照してください。
+このページは 3 つの基本 MiniMax H3 テンプレートとネイティブノードの高度なテクニックを扱います。機能とライセンスは [MiniMax H3 概要](/ja/tutorials/video/minimax/minimax-h3) を参照してください。
 
 ## MiniMax H3 テキストからビデオ (T2V)
 
@@ -74,7 +74,7 @@ ComfyUI/
 3. **解像度**: 16:9 ネイティブ **1344×768**（短辺 768px）; **32** 倍数
 4. **長さ**: 24fps で **17k+5** グリッド
 5. **先頭/末尾フレーム**: `MiniMaxH3ImageToVideo` の `first_frame` / `last_frame` で FL2VA
-6. **Turbo**: 既定 **20** ステップ; `turbo_mode` で **8** ステップ Lightning LoRA（速い、モーション/音声ややソフト）。モーション品質は **25** などに増やす
+6. **Turbo**: 既定 **20** ステップ; `turbo_mode` で **8** ステップ Lightning LoRA（速い、モーション/音声ややソフト）。モーション品質は **25** などに増やす。速い動作とカメラワークは短い turbo スケジュールで最初に甘くなります。静止カットではなく実際に動くクリップでステップ数を比べてください
 
 公式ベースガイド: [プロンプトガイド](/ja/tutorials/video/minimax/minimax-h3-prompt-guide)。
 
@@ -145,7 +145,7 @@ ComfyUI/
 1. **キーフレーム**: `first_frame` / `last_frame` は任意; その間のモーションを生成
 2. **プロンプト**: ショット、モーション、音声を 1 ブロックに
 3. **解像度/長さ**: ルールは T2V と同じ（ネイティブ目標 **1344×768**、**17k+5**）
-4. **Turbo**: `turbo_mode` は `turbo_steps`（既定 **8**）と `turbo_model_strength`（既定 **1.0**）で 20 ステップの代わり
+4. **Turbo**: `turbo_mode` は `turbo_steps`（既定 **8**）と `turbo_model_strength`（既定 **1.0**）で 20 ステップの代わり。短い turbo スケジュールでは速い動作とカメラワークが最初に甘くなるため、turbo を決める前に確認してください
 
 `first_frame` は縦横比を保たずにキャンバスへ拡縮されるため、アスペクト比の違う画像は引き伸ばされます。`last_frame` は中央でクロップされます。I2V テンプレートは **1:1**、**0.4** MP（**640×640**）で開くので、入力画像に合わせてアスペクト比を設定し、最終出力の前に Megapixels を上げてください。
 
@@ -258,3 +258,13 @@ H3 ノードの `positive` と latent を `MiniMaxH3AddGuide` に接続:
 [トークン単位マスク](https://github.com/Comfy-Org/ComfyUI/pull/15375) で動画・音声 latent の一部を再生成。
 
 サンプラー `denoise_mask` にマスク: **0** 保持、**1** 再生成。動画は 2×2 latent パッチ、音声は latent フレーム単位。インペイント、削除、拡張に使用。
+
+### 音声を先に作る
+
+音声と動画は同じデノイズで生成されるため、短いスケジュールや turbo LoRA は映像と一緒に音声も甘くします。2 パスに分ければ切り離せます。まず小さいキャンバスで音声を固め、その音声を条件に映像を作ります。
+
+1. **安い音声パス**: **32×32** で `Boolean (Enable Lightning LoRA)` をオフ、**40–50** ステップ、最終カットに必要な長さでテキストから動画のパスを 1 回実行します。音声 latent のフレーム数はクリップ長に従い、キャンバスサイズには依存しないため、このパスはフル解像度のレンダーよりはるかに安いコストでフル尺の音声を作れます。Resolution Selector を切り離し、対応する `width` と `height` にサイズを直接入力します。セレクターの下限は **0.1** メガピクセルです。ステップ数はサブグラフ内で **20** に設定された `PrimitiveInt` にあります（**6** のほうは turbo のステップ数です）。このパスでは音声だけを使います。
+2. **トラックを保存**: サブグラフは内部で音声をデコードし、映像だけを出力します。サブグラフ内に `SaveAudioAdvanced` を追加し、`VAEDecodeAudio` を接続してください。
+3. **2 パス目を条件づける**: R2V ワークフローに `LoadAudio` を追加して `ref_audio` に接続し、`audio_vae` は接続したままにします。プロンプトでは `<Audio 1>` として参照し、関係を明記します。`reference` は信号をコピーせず音色・リズム・話し方だけを参照し、`fully_copy` は完全な最終トラックとして扱い、`partially_copy` はタイムラインの一部または一部の音声レイヤーだけをコピーし、`weak_reference` はカテゴリや雰囲気の広い類似だけを保ちます。2 パス目はゼロから音声を作るのではなく、確定した声とタイミングから始まります。まったく同じファイルを使う場合は、プロンプトに頼らず `LoadAudio` を `CreateVideo` に接続します。
+
+キャンバスサイズは音色とレンダー時間の交換になります。32×32 のパスはほぼ無料ですが、そのサイズでは声に機械的な響きが乗ることがあります。声を最終カットで保たせたいときはキャンバスを上げてください。そのぶんパスのコストも上がります。台詞や歌詞は小さいキャンバスでも概ね保ちます。画面の動きに合わせた効果音やスコアは保たないので、2 パス目で生成し、`partially_copy` で 1 パス目のトラックと結合します。
diff --git a/ko/tutorials/video/ltx/ltx-2-5.mdx b/ko/tutorials/video/ltx/ltx-2-5.mdx
index 00cf27a7bf..e459af3f62 100644
--- a/ko/tutorials/video/ltx/ltx-2-5.mdx
+++ b/ko/tutorials/video/ltx/ltx-2-5.mdx
@@ -2,12 +2,12 @@
 title: "LTX-2.5: ComfyUI 비디오·오디오·멀티샷 워크플로"
 description: "Lightricks의 오픈 웨이트 비디오 모델을 ComfyUI에서 사용하세요. 텍스트→비디오, 이미지→비디오, 첫/마지막 프레임, DFR 워크플로와 멀티샷 장면, 오디오 동기화, 4K HDR을 지원합니다."
 sidebarTitle: "LTX-2.5"
-translationSourceHash: 4a7cea51
+translationSourceHash: ebef3f6f
 translationFrom: tutorials/video/ltx/ltx-2-5.mdx
 translationBlockHashes:
   "_intro": 4aa7d7de
   "Key features": 5bfa64be
-  "Getting started": b7968eb6
+  "Getting started": 7909413a
   "ComfyUI Native Workflows": a65c6f64
   "LTX-2.5 DFR: Image to Video": 3899f68b
   "Resources": f89b58b6
@@ -50,6 +50,10 @@ LTX-2.5는 ComfyUI에서 기본적으로 지원됩니다. 시작하려면:
   LTX-2.5 모델 파일은 Hugging Face의 액세스 제한(gated) 저장소인 [Lightricks/LTX-2.5](https://huggingface.co/Lightricks/LTX-2.5)와 [Lightricks/LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler](https://huggingface.co/Lightricks/LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler)에 호스팅되어 있습니다. 저장소 페이지를 열고 모델 라이선스에 동의한 후 액세스 요청이 승인될 때까지 기다렸다가 모델을 다운로드하세요. 액세스 권한이 없으면 모델 다운로드가 실패합니다.
 </Callout>
 
+<Note>
+LTX-2.5에는 같은 latent 공간을 쓰는 비디오 VAE가 두 개 있습니다. 템플릿은 확산 디코더인 [ltx-2.5-video-vae-bf16.safetensors](https://huggingface.co/Lightricks/LTX-2.5/blob/main/vae/ltx-2.5-video-vae-bf16.safetensors)를 불러옵니다. 학습된 디코드 단계를 거치고 오프로드할 수 없으므로 전체 크기 디코드가 많은 VRAM을 차지합니다. 그 디코드에서 메모리가 부족하면 `VAELoader`를 [ltx-2.5-video-vae-conv-bf16.safetensors](https://huggingface.co/Lightricks/LTX-2.5/blob/main/vae/ltx-2.5-video-vae-conv-bf16.safetensors)로 바꾸세요. 이 파일은 합성곱 오토인코더로 디코드하므로 더 빠르고 가벼우며, 얼굴과 화면 속 글자는 조금 부드러워집니다.
+</Note>
+
 ## ComfyUI 네이티브 워크플로
 
 ### LTX-2.5 텍스트 기반 비디오 생성 (T2V)
diff --git a/ko/tutorials/video/minimax/minimax-h3-native.mdx b/ko/tutorials/video/minimax/minimax-h3-native.mdx
index 27da4717e1..ab26e4077f 100644
--- a/ko/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/ko/tutorials/video/minimax/minimax-h3-native.mdx
@@ -2,17 +2,17 @@
 title: "ComfyUI MiniMax H3 텍스트·이미지·레퍼런스 기반 비디오 워크플로"
 description: "ComfyUI의 MiniMax H3 네이티브 워크플로 단계별 예제입니다. 가이드 고정과 잠재 노이즈 마스크로 생성 비디오를 더 세밀하게 제어할 수 있습니다."
 sidebarTitle: "네이티브 워크플로"
-translationSourceHash: 9e8d385c
+translationSourceHash: f8176106
 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx
 translationBlockHashes:
-  "_intro": 4f5274f5
-  "MiniMax H3 Text to Video (T2V)": 8807c872
-  "MiniMax H3 Image to Video (I2V)": c59554b0
+  "_intro": dde651e5
+  "MiniMax H3 Text to Video (T2V)": 42ec8cfb
+  "MiniMax H3 Image to Video (I2V)": 2a06e3fa
   "MiniMax H3 Reference to Video (R2V)": 961829df
-  "Advanced workflows with the native nodes": 37164e51
+  "Advanced workflows with the native nodes": 85460bff
 ---
 
-이 페이지는 세 가지 기본 MiniMax H3 템플릿과 네이티브 노드 고급 기법을 다룹니다. 기능, 튜닝, 가속은 [MiniMax H3 개요](/ko/tutorials/video/minimax/minimax-h3)를 참조하세요.
+이 페이지는 세 가지 기본 MiniMax H3 템플릿과 네이티브 노드 고급 기법을 다룹니다. 기능과 라이선스는 [MiniMax H3 개요](/ko/tutorials/video/minimax/minimax-h3)를 참조하세요.
 
 ## MiniMax H3 텍스트 기반 비디오 생성 (T2V)
 
@@ -74,7 +74,7 @@ ComfyUI/
 3. **해상도**: 16:9 네이티브 **1344×768**(단변 768px); **32** 배수
 4. **길이**: 24fps **17k+5** 그리드
 5. **첫/끝 프레임**: `MiniMaxH3ImageToVideo` `first_frame`/`last_frame`으로 FL2VA
-6. **Turbo**: 기본 **20** 스텝; `turbo_mode` **8** 스텝 Lightning LoRA(빠름, 모션/오디오 약간 soft). 모션 품질은 **25** 등
+6. **Turbo**: 기본 **20** 스텝; `turbo_mode` **8** 스텝 Lightning LoRA(빠름, 모션/오디오 약간 soft). 모션 품질은 **25** 등. 빠른 동작과 카메라 무브는 짧은 turbo 스케줄에서 가장 먼저 뭉개집니다. 정지 샷이 아니라 실제 움직임이 있는 클립에서 스텝 수를 비교하세요
 
 공식 기본 가이드: [프롬프트 가이드](/ko/tutorials/video/minimax/minimax-h3-prompt-guide).
 
@@ -145,7 +145,7 @@ ComfyUI/
 1. **키프레임**: `first_frame`/`last_frame` 선택; 그 사이 모션 생성
 2. **프롬프트**: 샷, 모션, 오디오 한 블록
 3. **해상도/길이**: 규칙은 T2V와 동일(네이티브 목표 **1344×768**, **17k+5**)
-4. **Turbo**: `turbo_mode`는 `turbo_steps`(기본 **8**), `turbo_model_strength`(기본 **1.0**)로 20 스텝 대체
+4. **Turbo**: `turbo_mode`는 `turbo_steps`(기본 **8**), `turbo_model_strength`(기본 **1.0**)로 20 스텝 대체. 짧은 turbo 스케줄에서는 빠른 동작과 카메라 무브가 가장 먼저 뭉개지므로 turbo를 확정하기 전에 확인하세요
 
 `first_frame`은 종횡비를 유지하지 않고 캔버스 크기에 맞춰 조정되므로 비율이 다른 이미지는 늘어납니다. `last_frame`은 중앙에서 크롭됩니다. I2V 템플릿은 **1:1**, **0.4** MP(**640×640**)로 열리므로 입력 이미지에 맞춰 종횡비를 설정하고, 최종 렌더 전에 Megapixels를 올리세요.
 
@@ -258,3 +258,13 @@ H3 노드 `positive`·latent를 `MiniMaxH3AddGuide`에 연결:
 [토큰 단위 마스크](https://github.com/Comfy-Org/ComfyUI/pull/15375)로 비디오·오디오 latent 일부 재생성.
 
 샘플러 `denoise_mask`: **0** 유지, **1** 재생성. 비디오 2×2 latent 패치, 오디오 latent 프레임 단위. 인페인팅, 제거, 확장.
+
+### 사운드트랙을 먼저 만들기
+
+오디오와 비디오는 같은 디노이즈에서 함께 생성되므로, 짧은 스케줄이나 turbo LoRA는 영상과 함께 오디오도 무디게 만듭니다. 두 번의 패스로 분리할 수 있습니다. 작은 캔버스에서 오디오를 먼저 확정하고, 그 오디오를 조건으로 영상을 만듭니다.
+
+1. **저렴한 오디오 패스**: **32×32**, `Boolean (Enable Lightning LoRA)` 끔, **40–50** 스텝, 최종 샷에 필요한 길이로 텍스트-투-비디오 패스를 한 번 실행합니다. 오디오 latent 프레임 수는 클립 길이를 따르고 캔버스 크기와 무관하므로, 이 패스는 전체 해상도 렌더보다 훨씬 낮은 비용으로 전체 길이 트랙을 만듭니다. Resolution Selector 연결을 끊고 해당 `width`, `height`에 값을 직접 입력하세요. 선택기의 하한은 **0.1** 메가픽셀입니다. 스텝 수는 서브그래프 안에서 **20**으로 설정된 `PrimitiveInt`에 있습니다(**6**인 쪽은 turbo 스텝 수입니다). 이 패스에서는 오디오만 사용합니다.
+2. **트랙 저장**: 서브그래프는 내부에서 오디오를 디코드하고 비디오만 노출하므로, 서브그래프 안에 `SaveAudioAdvanced`를 추가하고 `VAEDecodeAudio`를 연결하세요.
+3. **두 번째 패스 조건 걸기**: R2V 워크플로에 `LoadAudio`를 추가해 `ref_audio`에 연결하고 `audio_vae`는 연결된 상태로 둡니다. 프롬프트에서 `<Audio 1>`로 참조하고 관계를 명시하세요. `reference`는 신호를 복사하지 않고 음색, 리듬, 말투만 참조하고, `fully_copy`는 완전한 최종 트랙으로 취급하며, `partially_copy`는 타임라인의 일부나 일부 오디오 레이어만 복사하고, `weak_reference`는 범주와 분위기의 넓은 유사성만 유지합니다. 두 번째 패스는 빈 상태에서 음성을 지어내는 대신 확정된 목소리와 타이밍에서 시작합니다. 완전히 같은 파일을 쓰려면 프롬프트에 의존하지 말고 `LoadAudio`를 `CreateVideo`에 연결하세요.
+
+캔버스 크기는 음질과 렌더 시간을 맞바꿉니다. 32×32 패스는 거의 무료지만, 그 크기에서는 목소리에 기계적인 느낌이 붙을 수 있습니다. 목소리가 최종 컷에서 버텨야 한다면 캔버스를 올리세요. 그만큼 패스 비용도 올라갑니다. 대사와 가사는 작은 캔버스에서도 대체로 버팁니다. 화면 동작에 맞춘 효과음과 스코어는 버티지 못하므로 두 번째 패스에서 만들고 `partially_copy`로 첫 트랙과 합치세요.
diff --git a/tutorials/video/ltx/ltx-2-5.mdx b/tutorials/video/ltx/ltx-2-5.mdx
index b4c65ea05c..5b541383b5 100644
--- a/tutorials/video/ltx/ltx-2-5.mdx
+++ b/tutorials/video/ltx/ltx-2-5.mdx
@@ -41,6 +41,10 @@ LTX-2.5 is natively supported in ComfyUI. To get started:
   The LTX-2.5 model files are hosted in the gated [Lightricks/LTX-2.5](https://huggingface.co/Lightricks/LTX-2.5) and [Lightricks/LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler](https://huggingface.co/Lightricks/LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler) repositories on Hugging Face. Open the repository pages, accept the model license, and wait for your access request to be approved before downloading the models. Model downloads will fail without access.
 </Callout>
 
+<Note>
+LTX-2.5 ships two video VAEs that share the same latent space. The templates load [ltx-2.5-video-vae-bf16.safetensors](https://huggingface.co/Lightricks/LTX-2.5/blob/main/vae/ltx-2.5-video-vae-bf16.safetensors), the diffusion decoder. It runs a learned decode step and cannot be offloaded, so a full-size decode holds a lot of VRAM. When that decode runs out of memory, switch `VAELoader` to [ltx-2.5-video-vae-conv-bf16.safetensors](https://huggingface.co/Lightricks/LTX-2.5/blob/main/vae/ltx-2.5-video-vae-conv-bf16.safetensors). The conv file decodes through the convolutional autoencoder, which is faster and lighter, with softer faces and on-screen text.
+</Note>
+
 ## ComfyUI Native Workflows
 
 ### LTX-2.5 Text to Video (T2V)
diff --git a/tutorials/video/minimax/minimax-h3-native.mdx b/tutorials/video/minimax/minimax-h3-native.mdx
index ac0d6102db..1af9f4e235 100644
--- a/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/tutorials/video/minimax/minimax-h3-native.mdx
@@ -4,7 +4,7 @@ description: "Step-by-step MiniMax H3 native workflow examples in ComfyUI, inclu
 sidebarTitle: "Native workflows"
 ---
 
-This page covers the three base MiniMax H3 templates plus advanced native-node techniques. For capabilities, tuning, and speedups, see the [MiniMax H3 overview](/tutorials/video/minimax/minimax-h3).
+This page covers the three base MiniMax H3 templates plus advanced native-node techniques. For capabilities and licensing, see the [MiniMax H3 overview](/tutorials/video/minimax/minimax-h3).
 
 ## MiniMax H3 Text to Video (T2V)
 
@@ -66,7 +66,7 @@ ComfyUI/
 3. **Resolution**: Native 16:9 is **1344×768** (768px short edge); round to multiples of **32**
 4. **Duration**: Snaps to **17k+5** frames at 24fps
 5. **First/last frame**: Wire images to `first_frame` / `last_frame` on `MiniMaxH3ImageToVideo` for FL2VA
-6. **Turbo**: Default **20** steps; enable `turbo_mode` for **8**-step Lightning LoRA (faster, slightly softer motion/audio). Raise steps (e.g. **25**) for motion quality
+6. **Turbo**: Default **20** steps; enable `turbo_mode` for **8**-step Lightning LoRA (faster, slightly softer motion/audio). Raise steps (e.g. **25**) for motion quality. Fast action and camera moves are the first thing a short turbo schedule softens. Compare step counts on a clip with real movement, not on a static shot
 
 Official base guide: [prompt guide](/tutorials/video/minimax/minimax-h3-prompt-guide).
 
@@ -137,7 +137,7 @@ ComfyUI/
 1. **Keyframes**: Optional `first_frame` / `last_frame`; motion is generated between them
 2. **Prompt**: Shots, motion, and audio in one block
 3. **Resolution / duration**: the same rules as T2V (**1344×768** native target, **17k+5** grid)
-4. **Turbo**: `turbo_mode` uses `turbo_steps` (default **8**) and `turbo_model_strength` (default **1.0**) instead of 20 steps
+4. **Turbo**: `turbo_mode` uses `turbo_steps` (default **8**) and `turbo_model_strength` (default **1.0**) instead of 20 steps. Fast action and camera moves soften first under a short turbo schedule, so check them before committing to turbo
 
 `first_frame` is resized to the canvas without keeping its aspect ratio, so an off-aspect image comes out stretched. `last_frame` is center-cropped instead. The I2V template opens at **1:1** and **0.4** MP (**640×640**), so set the aspect ratio to your image and raise Megapixels for a final render.
 
@@ -250,3 +250,13 @@ See the [Multiframe Reference](/tutorials/video/minimax/minimax-h3-multiframe) t
 [Per-token noise masks](https://github.com/Comfy-Org/ComfyUI/pull/15375) on video and audio latents let you regenerate part of a clip.
 
 Connect a mask to the sampler `denoise_mask`: **0** = keep, **1** = regenerate. Video masks snap to the 2×2 latent patch grid; audio masks to whole latent frames. Use for inpainting, removal, or extension while holding the rest stable.
+
+### Generating the soundtrack first
+
+Audio and video are denoised together, so a short schedule or a turbo LoRA softens the track along with the picture. Two passes separate them: settle the audio on a small canvas, then condition the video on it.
+
+1. **Cheap audio pass**: Run a text-to-video pass at **32×32** with `Boolean (Enable Lightning LoRA)` off, **40–50** steps, and the duration the final shot needs. Audio latent frames follow clip length, not canvas size, so this pass produces a full-length track for far less than a full-resolution render. Unlink the Resolution Selector and type the size into its `width` and `height` targets; the selector bottoms out at **0.1** Megapixels. Steps live on the `PrimitiveInt` set to **20** inside the subgraph (the one set to **6** is the turbo step count). Only the audio is kept.
+2. **Save the track**: The subgraph decodes audio inside and exposes only the video, so add a `SaveAudioAdvanced` node inside the subgraph and connect `VAEDecodeAudio` to it.
+3. **Condition the second pass**: In the R2V workflow, add `LoadAudio`, wire it to `ref_audio`, and keep `audio_vae` connected. Cite the track as `<Audio 1>` and state the relationship: `reference` keeps the timbre, rhythm, and delivery without copying the signal, `fully_copy` makes it the complete final track, `partially_copy` keeps part of the timeline or selected layers, and `weak_reference` keeps only broad similarity. The second pass starts from a settled voice and timing instead of inventing speech from nothing. For the identical file, wire `LoadAudio` into `CreateVideo` instead of relying on the prompt.
+
+Canvas size trades voice quality for render time. A 32×32 pass is nearly free, but voices can pick up a robotic edge at that size. Raise the canvas when the voice has to hold up in the final cut, and expect the pass to cost more as you do. Speech and lyrics usually survive the small canvas. Sound effects and scoring tied to on-screen action do not, so render those in the second pass and let `partially_copy` combine them with the first track.
diff --git a/zh/tutorials/video/ltx/ltx-2-5.mdx b/zh/tutorials/video/ltx/ltx-2-5.mdx
index 398d80cf84..3bdec5f3d0 100644
--- a/zh/tutorials/video/ltx/ltx-2-5.mdx
+++ b/zh/tutorials/video/ltx/ltx-2-5.mdx
@@ -2,12 +2,12 @@
 title: "LTX-2.5：ComfyUI 视频、音频与多镜头工作流"
 description: "Lightricks 开源视频模型在 ComfyUI 的原生工作流：文生视频、图生视频、首尾帧生视频与 DFR 工作流，支持多镜头场景、音画同步和 4K HDR。"
 sidebarTitle: "LTX-2.5"
-translationSourceHash: 4a7cea51
+translationSourceHash: ebef3f6f
 translationFrom: tutorials/video/ltx/ltx-2-5.mdx
 translationBlockHashes:
   "_intro": 4aa7d7de
   "Key features": 5bfa64be
-  "Getting started": b7968eb6
+  "Getting started": 7909413a
   "ComfyUI Native Workflows": a65c6f64
   "LTX-2.5 DFR: Image to Video": 3899f68b
   "Resources": f89b58b6
@@ -49,6 +49,10 @@ ComfyUI 原生支持 LTX-2.5。开始使用：
   LTX-2.5 模型文件托管在 Hugging Face 上受控访问（gated）的 [Lightricks/LTX-2.5](https://huggingface.co/Lightricks/LTX-2.5) 和 [Lightricks/LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler](https://huggingface.co/Lightricks/LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler) 仓库中。打开仓库页面，接受模型许可协议，并等待访问申请获批后再下载模型。未获得访问权限时，模型下载会失败。
 </Callout>
 
+<Note>
+LTX-2.5 有两种视频 VAE，共用同一套 latent 空间。模板加载 [ltx-2.5-video-vae-bf16.safetensors](https://huggingface.co/Lightricks/LTX-2.5/blob/main/vae/ltx-2.5-video-vae-bf16.safetensors)，即扩散解码器。它多跑一段学习式解码，且无法卸载，全尺寸解码会占住大量显存。解码显存不够时，把 `VAELoader` 换成 [ltx-2.5-video-vae-conv-bf16.safetensors](https://huggingface.co/Lightricks/LTX-2.5/blob/main/vae/ltx-2.5-video-vae-conv-bf16.safetensors)。后者走卷积自编码器解码，更快更轻，人脸与画面文字会软一些。
+</Note>
+
 ## ComfyUI 原生工作流
 
 ### LTX-2.5 文生视频 (T2V)
diff --git a/zh/tutorials/video/minimax/minimax-h3-native.mdx b/zh/tutorials/video/minimax/minimax-h3-native.mdx
index 2162165410..3ca6b014ff 100644
--- a/zh/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/zh/tutorials/video/minimax/minimax-h3-native.mdx
@@ -2,16 +2,16 @@
 title: "ComfyUI MiniMax H3 文生视频、图生视频与参考生视频工作流"
 description: "ComfyUI 中 MiniMax H3 原生工作流的逐步示例，含帧锚定与潜在噪声蒙版技巧，帮助你对生成视频做更精细的控制。"
 sidebarTitle: "原生工作流"
-translationSourceHash: 9e8d385c
+translationSourceHash: f8176106
 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx
 translationBlockHashes:
-  "_intro": 4f5274f5
-  "MiniMax H3 Text to Video (T2V)": 8807c872
-  "MiniMax H3 Image to Video (I2V)": c59554b0
+  "_intro": dde651e5
+  "MiniMax H3 Text to Video (T2V)": 42ec8cfb
+  "MiniMax H3 Image to Video (I2V)": 2a06e3fa
   "MiniMax H3 Reference to Video (R2V)": 961829df
-  "Advanced workflows with the native nodes": 37164e51
+  "Advanced workflows with the native nodes": 85460bff
 ---
-本页涵盖三个基础 MiniMax H3 模板及原生节点高级技巧。能力、调参与加速见 [MiniMax H3 概览](/zh/tutorials/video/minimax/minimax-h3)。
+本页涵盖三个基础 MiniMax H3 模板及原生节点高级技巧。能力与授权见 [MiniMax H3 概览](/zh/tutorials/video/minimax/minimax-h3)。
 
 ## MiniMax H3 文生视频（T2V）
 
@@ -73,7 +73,7 @@ ComfyUI/
 3. **分辨率**：16:9 原生 **1344×768**（短边 768px）；对齐 **32** 倍数
 4. **时长**：24fps 下对齐 **17k+5** 帧网格
 5. **首尾帧**：在 `MiniMaxH3ImageToVideo` 的 `first_frame`/`last_frame` 接图即为 FL2VA
-6. **Turbo**：默认 **20** 步；`turbo_mode` 为 **8** 步 Lightning LoRA（更快，运动/音频略软）。可提高步数（如 **25**）改善运动
+6. **Turbo**：默认 **20** 步；`turbo_mode` 为 **8** 步 Lightning LoRA（更快，运动/音频略软）。可提高步数（如 **25**）改善运动。快速动作与镜头运动最先被短 turbo 调度抹平。要在有真实运动的片段上比较步数，别用静止镜头判断
 
 官方基础指南：[提示词指南](/zh/tutorials/video/minimax/minimax-h3-prompt-guide)。
 
@@ -144,7 +144,7 @@ ComfyUI/
 1. **关键帧**：`first_frame`/`last_frame` 可选，运动生成在两者之间
 2. **提示词**：镜头、运动、音频写在同一块
 3. **分辨率/时长**：规则同 T2V（原生目标 **1344×768**，**17k+5**）
-4. **Turbo**：`turbo_mode` 使用 `turbo_steps`（默认 **8**）与 `turbo_model_strength`（默认 **1.0**），替代 20 步
+4. **Turbo**：`turbo_mode` 使用 `turbo_steps`（默认 **8**）与 `turbo_model_strength`（默认 **1.0**），替代 20 步。短 turbo 调度下，快速动作与镜头运动最先变软，定下 turbo 前先看这两项
 
 `first_frame` 会被缩放到画布，不保持原比例，所以宽高比不一致的图会被拉伸；`last_frame` 则按中心裁剪。I2V 模板默认 **1:1**、**0.4** MP（**640×640**），请把宽高比设成与输入图一致，正式出片前再调高 Megapixels。
 
@@ -257,3 +257,13 @@ ComfyUI/
 [逐 token 遮罩](https://github.com/Comfy-Org/ComfyUI/pull/15375) 覆盖视频与音频 latent，可只重生成一部分。
 
 遮罩接采样器 `denoise_mask`：**0** 保留，**1** 重生成。视频对齐 2×2 latent 块，音频对齐整帧。用于修补、移除或扩展。
+
+### 先生成音轨
+
+音频与视频在同一次去噪中生成，短调度或 turbo LoRA 会连同画面一起把音轨变软。分两遍就能拆开：先在小画布上把音频定下来，再让视频以它为条件。
+
+1. **便宜的音频遍**：以 **32×32** 画布、关闭 `Boolean (Enable Lightning LoRA)`、**40–50** 步、按最终镜头需要的时长跑一遍文生视频。音频 latent 的帧数跟着片段长度走，与画布尺寸无关，所以这一遍能用远低于一次全分辨率渲染的代价产出完整长度的音轨。断开 Resolution Selector，把尺寸直接填进它对应的 `width` 和 `height`；选择器最小只到 **0.1** 百万像素。步数在子图里设为 **20** 的那个 `PrimitiveInt` 上（设为 **6** 的那个是 turbo 步数）。这一遍只保留音频。
+2. **存下音轨**：子图在内部解码音频，只把视频暴露出来，所以在子图内加一个 `SaveAudioAdvanced`，把 `VAEDecodeAudio` 接到它上面。
+3. **给第二遍做条件**：在 R2V 工作流中加入 `LoadAudio`，接到 `ref_audio`，并保持 `audio_vae` 连接。提示词里用 `<Audio 1>` 引用它，并写明关系：`reference` 只取音色、节奏与语气，不复制信号本身；`fully_copy` 把它当作完整的最终音轨；`partially_copy` 只复制一部分时间线或部分音轨层；`weak_reference` 只保留大类与氛围上的相似。第二遍于是从一个已经确定的音色与节奏起步，而不是从零编造语音。要用完全相同的文件，把 `LoadAudio` 接到 `CreateVideo`，不要依赖提示词。
+
+画布尺寸是拿音色换渲染时间。32×32 的一遍几乎免费，但这个尺寸下声音可能带上机械感。音色要在成片里撑得住，就把画布调大，代价是这个遍数变贵。语音与歌词通常能扛住小画布；跟着画面动作走的音效与配乐扛不住，这类内容留给第二遍生成，再用 `partially_copy` 与第一遍的音轨合并。

From 9db77372e2879338cd7e8d3c76c46edd4ef6613c Mon Sep 17 00:00:00 2001
From: lin-bot23 <linzhixiong23@gmail.com>
Date: Fri, 30 Oct 2026 18:07:50 +0800
Subject: [PATCH 07/13] docs(minimax): cover dialogue line placement and
 punctuation on the H3 prompt guide

---
 .../video/minimax/minimax-h3-prompt-guide.mdx        | 10 ++++++----
 .../video/minimax/minimax-h3-prompt-guide.mdx        | 12 +++++++-----
 .../video/minimax/minimax-h3-prompt-guide.mdx        | 12 +++++++-----
 tutorials/video/minimax/minimax-h3-prompt-guide.mdx  |  4 +++-
 .../video/minimax/minimax-h3-prompt-guide.mdx        | 12 +++++++-----
 5 files changed, 30 insertions(+), 20 deletions(-)

diff --git a/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
index 34416cf910..ae6af748be 100644
--- a/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
+++ b/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
@@ -2,14 +2,14 @@
 title: "Guía de prompts de ComfyUI MiniMax H3: guías oficiales, consejos y embeddings de prompt"
 description: "Escribe mejores prompts de MiniMax H3 en ComfyUI: guías oficiales, consejos generales, reglas de diálogo y hablantes, y embeddings de estilo."
 sidebarTitle: "Guía de prompts"
-translationSourceHash: 0d517d21
+translationSourceHash: a2bc4fff
 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx
 translationBlockHashes:
   "_intro": da22cc9d
   "Official guides": 3d762c67
-  "General tips": 91822299
+  "General tips": fab4ad35
   "Shots and cuts": 8fc80fd4
-  "Dialogue and speakers": 4985438e
+  "Dialogue and speakers": e403c931
   "Prompt embeddings": 9c899169
 ---
 H3 espera planos, cámara, diálogo y SFX en un solo prompt multimodal. Esta página reúne guías oficiales, consejos entre flujos de trabajo, reglas de planos y diálogo, y embeddings de estilo.
@@ -86,7 +86,9 @@ Escribe el movimiento dentro de la acción. No acumules las palabras como etique
 
 - **IDs de hablante**: `(S1)`, `(S2)`, … para cada voz (incluidas las que están fuera de pantalla y las que cantan). Los IDs indexan hablantes, no números de personaje. Dúo: `(S1,S2)`
 - **Contenido de `<d>`**: el ID, la acción y la forma de decirlo van fuera de las etiquetas; dentro de las etiquetas, la etiqueta de idioma + las palabras pronunciadas textualmente, en el párrafo del plano donde se dicen
-- **Puntuación en `<d>`**: conserva `, . ? !`; elimina repeticiones decorativas y emoji; termina las afirmaciones con `.` `?` `!` antes de `</d>`; usa `[unclear]` cuando no estés seguro. Los puntos suspensivos suelen perjudicar la inteligibilidad
+- **Puntuación en `<d>`**: conserva `, . ? !`; elimina repeticiones decorativas y emoji; usa `[unclear]` cuando no estés seguro. Los puntos suspensivos suelen perjudicar la inteligibilidad. Cierra cada línea con un solo signo final antes de `</d>`, sin signos repetidos (`?!?`) y sin nada después de la etiqueta (`ready?</d>.`)
+- **Análisis de retención (R2V)**: describe qué conserva cada referencia, no lo que se dice. No escribas diálogo ni letras ahí, ni IDs `(Sx)`. Cada línea hablada va solo dentro de `<d>` en el plano donde se oye
+- **Referencias solo de timbre**: cuando `<Audio N>` aporta solo timbre, ritmo, emoción o forma de decirlo, no reutilices sus palabras. Escribe la línea nueva en `<d>` y marca el audio como `reference`, por ejemplo `<Audio 2>: reference - the target speaker follows <Audio 2>'s voice timbre and measured delivery without copying the original signal`
 - **Voz en off**: usa `says in an off-screen voiceover` e indica que los labios permanecen cerrados después de `</d>`. Diálogo con corte cruzado: `<scenetrans>` en ambos lados y el audio continúa
 - **Varios hablantes + referencias de audio**: puede ocurrir una asignación incorrecta del hablante. Vincula las líneas a eventos visibles, no a códigos de tiempo. Si hace falta, genera el audio de cada hablante por separado de forma externa y referéncialo con el marcador correcto (`fully_copy` vs `partially_copy`)
 
diff --git a/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
index 5c210d85ad..bd8ca08c28 100644
--- a/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
+++ b/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
@@ -2,14 +2,14 @@
 title: "ComfyUI MiniMax H3 プロンプトガイド: 公式ガイド、ヒント、プロンプト埋め込み"
 description: "ComfyUI でより良い MiniMax H3 プロンプトを書く方法: MiniMax の公式プロンプト作成ガイド、一般的なプロンプトのコツ、セリフと話者のルール、スタイル埋め込み。"
 sidebarTitle: "プロンプトガイド"
-translationSourceHash: 0d517d21
+translationSourceHash: a2bc4fff
 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx
 translationBlockHashes:
   "_intro": da22cc9d
   "Official guides": 3d762c67
-  "General tips": 91822299
+  "General tips": fab4ad35
   "Shots and cuts": 8fc80fd4
-  "Dialogue and speakers": 4985438e
+  "Dialogue and speakers": e403c931
   "Prompt embeddings": 9c899169
 ---
 
@@ -87,9 +87,11 @@ non_diegetic_music: ...
 
 - **話者 ID**: `(S1)`、`(S2)`…（オフ画面/歌唱含む）; デュエット `(S1,S2)`
 - **`<d>`**: ID・動作・演技はタグ外; タグ内は言語タグと原文
-- **`<d>` 内句読点**: `, . ? !` を保持; 文末は `.` `?` `!`; 不明は `[unclear]`; 省略号は避ける
+- **`<d>` 内句読点**: `, . ? !` を保持; 装飾的な繰り返しや絵文字は削除; 不明は `[unclear]`; 省略号は避ける; 各行は `</d>` の前に文末記号を 1 つだけ置き、重ねた記号（`?!?`）や閉じタグの後ろの記号（`ready?</d>.`）は置かない
+- **retention_analysis（R2V）**: 参照が何を保持したかだけを書き、セリフや歌詞、`(Sx)` は置かない; セリフはそれが聞こえるショットの `<d>` 内にのみ置く
+- **音色のみの参照**: `<Audio N>` が音色・リズム・感情・演技だけを提供する場合はその言葉を使わない; 新しいセリフを `<d>` に書き、音声は `reference` とする（例: `<Audio 2>: reference - the target speaker follows <Audio 2>'s voice timbre and measured delivery without copying the original signal`）
 - **ボイスオーバー**: `says in an off-screen voiceover` と唇が閉じたまま; 跨ぎは `<scenetrans>`
-- **複数話者 + オーディオ参照**: 誤割当あり; タイムコードより可視イベントに結び付け; 必要なら話者別に外部生成して参照
+- **複数話者 + オーディオ参照**: 誤割当あり; タイムコードより可視イベントに結び付け; 必要なら話者別に外部生成し、用途に合うマーカー（`fully_copy` か `partially_copy`）で参照する
 
 ## プロンプト埋め込み
 
diff --git a/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
index af81b2dd39..8254f53be1 100644
--- a/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
+++ b/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
@@ -2,14 +2,14 @@
 title: "ComfyUI MiniMax H3 프롬프트 가이드: 공식 가이드, 팁, 프롬프트 임베딩"
 description: "ComfyUI에서 더 나은 MiniMax H3 프롬프트 작성하기: MiniMax 공식 프롬프트 작성 가이드, 일반 프롬프트 팁, 대사와 화자 규칙, 스타일 임베딩."
 sidebarTitle: "프롬프트 가이드"
-translationSourceHash: 0d517d21
+translationSourceHash: a2bc4fff
 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx
 translationBlockHashes:
   "_intro": da22cc9d
   "Official guides": 3d762c67
-  "General tips": 91822299
+  "General tips": fab4ad35
   "Shots and cuts": 8fc80fd4
-  "Dialogue and speakers": 4985438e
+  "Dialogue and speakers": e403c931
   "Prompt embeddings": 9c899169
 ---
 
@@ -87,9 +87,11 @@ non_diegetic_music: ...
 
 - **화자 ID**: `(S1)`, `(S2)`…(오프스크린/노래 포함); 듀엣 `(S1,S2)`
 - **`<d>`**: ID·동작·연기는 태그 밖; 태그 안 언어 태그+원문
-- **`<d>` 문장부호**: `, . ? !` 유지; 문장 끝 `.` `?` `!`; 불명 `[unclear]`; 말줄임 자제
+- **`<d>` 문장부호**: `, . ? !` 유지; 장식용 반복과 이모지 제거; 불명 `[unclear]`; 말줄임 자제; 각 줄은 `</d>` 앞에 문장 끝 부호를 하나만 두고, 겹친 부호(`?!?`)나 닫는 태그 뒤의 부호(`ready?</d>.`)는 쓰지 않는다
+- **retention_analysis(R2V)**: 참조가 무엇을 유지했는지만 적고 대사·가사·`(Sx)`는 넣지 않는다; 대사는 그것이 들리는 샷의 `<d>` 안에만 둔다
+- **음색만 참조**: `<Audio N>`이 음색·리듬·감정·전달만 제공할 때는 그 말을 그대로 쓰지 않는다; 새 대사를 `<d>`에 쓰고 오디오는 `reference`로 표시한다 (예: `<Audio 2>: reference - the target speaker follows <Audio 2>'s voice timbre and measured delivery without copying the original signal`)
 - **보이스오버**: `says in an off-screen voiceover` 및 입 닫힘; 넘김 `<scenetrans>`
-- **다화자+오디오 참조**: 오배정 가능; 타임코드보다 가시 이벤트; 필요 시 화자별 외부 생성 참조
+- **다화자+오디오 참조**: 오배정 가능; 타임코드보다 가시 이벤트; 필요 시 화자별로 외부 생성해 용도에 맞는 마커(`fully_copy` 또는 `partially_copy`)로 참조한다
 
 ## 프롬프트 임베딩
 
diff --git a/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
index 284159fb92..2c4b59cd6b 100644
--- a/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
+++ b/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
@@ -78,7 +78,9 @@ Write the move into the action. Do not stack the words as labels at the end of t
 
 - **Speaker IDs**: `(S1)`, `(S2)`, … for every voice (including off-screen/singing). IDs index speakers, not character numbers. Duet: `(S1,S2)`
 - **`<d>` content**: ID, action, and delivery outside tags; inside tags, language tag + spoken words verbatim, in the shot paragraph where spoken
-- **Punctuation in `<d>`**: Keep `, . ? !`; drop decorative repeats/emoji; end statements with `.` `?` `!` before `</d>`; use `[unclear]` when unsure. Ellipses often hurt intelligibility
+- **Punctuation in `<d>`**: Keep `, . ? !`; drop decorative repeats and emoji; use `[unclear]` when unsure. Ellipses often hurt intelligibility. Close each line with one terminal mark before `</d>`, with no repeated marks (`?!?`) and nothing after the tag (`ready?</d>.`)
+- **Retention analysis (R2V)**: It records what each reference keeps, not what is said. Write no dialogue or lyrics there and no `(Sx)` IDs. Every spoken line stays inside `<d>` in the shot where it is heard
+- **Timbre-only audio references**: When `<Audio N>` supplies only timbre, rhythm, emotion, or delivery, do not reuse its words. Write the new line in `<d>` and mark the audio `reference`, for example `<Audio 2>: reference - the target speaker follows <Audio 2>'s voice timbre and measured delivery without copying the original signal`
 - **Voiceover**: Use `says in an off-screen voiceover` and state lips stay closed after `</d>`. Cross-cut dialogue: `<scenetrans>` at both sides plus audio carries over
 - **Multi-speaker + audio refs**: Wrong-speaker assignment can happen. Tie lines to visible events, not timecodes. If needed, generate per-speaker audio externally and reference with the right marker (`fully_copy` vs `partially_copy`)
 
diff --git a/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
index ef8e632b50..1110eed759 100644
--- a/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
+++ b/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
@@ -2,14 +2,14 @@
 title: "ComfyUI MiniMax H3 提示词指南：官方指南、技巧与提示词嵌入"
 description: "在 ComfyUI 中写出更好的 MiniMax H3 提示词：MiniMax 官方提示词编写指南、通用提示词技巧、对白与说话人规则，以及风格嵌入。"
 sidebarTitle: "提示词指南"
-translationSourceHash: 0d517d21
+translationSourceHash: a2bc4fff
 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx
 translationBlockHashes:
   "_intro": da22cc9d
   "Official guides": 3d762c67
-  "General tips": 91822299
+  "General tips": fab4ad35
   "Shots and cuts": 8fc80fd4
-  "Dialogue and speakers": 4985438e
+  "Dialogue and speakers": e403c931
   "Prompt embeddings": 9c899169
 ---
 
@@ -87,9 +87,11 @@ non_diegetic_music: ...
 
 - **说话人 ID**：`(S1)`、`(S2)`…（含画外/演唱）；合唱 `(S1,S2)`
 - **`<d>`**：ID、动作、语气在标签外；标签内为语言标记与原文
-- **标点**：保留 `, . ? !`；句末用 `.` `?` `!`；听不清写 `[unclear]`；慎用省略号
+- **标点**：保留 `, . ? !`；去掉装饰性重复与 emoji；听不清写 `[unclear]`；慎用省略号；每句在 `</d>` 前以一个句末标点收尾，不连写（`?!?`），也不在结束标签后加标点（`ready?</d>.`）
+- **保留分析（R2V）**：只写参考保留了什么，不写台词；`retention_analysis` 中不放对白、歌词和 `(Sx)`；每句台词只放在听到它的那个镜头的 `<d>` 内
+- **仅音色参考**：`<Audio N>` 只提供音色、节奏、情绪或语气时不沿用其原话；新台词写进 `<d>`，音频标为 `reference`，例如 `<Audio 2>: reference - the target speaker follows <Audio 2>'s voice timbre and measured delivery without copying the original signal`
 - **画外音**：`says in an off-screen voiceover`，并说明嘴唇闭合；跨切用 `<scenetrans>`
-- **多说话人+音频参考**：易错配；绑定可见动作而非时间码；必要时外生成每人的音频参考
+- **多说话人+音频参考**：易错配；绑定可见动作而非时间码；必要时外部生成每人的音频参考，并按用法选标记（`fully_copy` 还是 `partially_copy`）
 
 ## 提示词嵌入
 

From 9079a7b043da7851b315b92c6d27b5d8a3c386c3 Mon Sep 17 00:00:00 2001
From: lin-bot23 <linzhixiong23@gmail.com>
Date: Sat, 31 Oct 2026 18:11:14 +0800
Subject: [PATCH 08/13] docs(minimax): document canvas size on the H3 prompt
 guide

---
 .../video/minimax/minimax-h3-prompt-guide.mdx | 23 ++++++++++++++++---
 .../video/minimax/minimax-h3-prompt-guide.mdx | 23 ++++++++++++++++---
 .../video/minimax/minimax-h3-prompt-guide.mdx | 23 ++++++++++++++++---
 .../video/minimax/minimax-h3-prompt-guide.mdx | 18 ++++++++++++++-
 .../video/minimax/minimax-h3-prompt-guide.mdx | 23 ++++++++++++++++---
 5 files changed, 97 insertions(+), 13 deletions(-)

diff --git a/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
index ae6af748be..ca36c6a60b 100644
--- a/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
+++ b/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
@@ -2,12 +2,13 @@
 title: "Guía de prompts de ComfyUI MiniMax H3: guías oficiales, consejos y embeddings de prompt"
 description: "Escribe mejores prompts de MiniMax H3 en ComfyUI: guías oficiales, consejos generales, reglas de diálogo y hablantes, y embeddings de estilo."
 sidebarTitle: "Guía de prompts"
-translationSourceHash: a2bc4fff
+translationSourceHash: cdaa2314
 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx
 translationBlockHashes:
   "_intro": da22cc9d
   "Official guides": 3d762c67
-  "General tips": fab4ad35
+  "General tips": 447ee59b
+  "Canvas size": 8c15c46c
   "Shots and cuts": 8fc80fd4
   "Dialogue and speakers": e403c931
   "Prompt embeddings": 9c899169
@@ -37,7 +38,7 @@ Escribe los prompts en **inglés**. Los diálogos y las letras de canciones van
 
 1. **Primero la escena**: la escena general, luego los planos cronometrados dentro de `[Shot 1]` de `integrated_multimodal_description`
 2. **Un solo bloque**: planos, cámara y audio juntos
-3. **Resolución**: nativa **1344×768** en 16:9; múltiplos de **32**
+3. **Resolución**: nativa **1344×768** en 16:9; múltiplos de **32**. Las plantillas abren por debajo de ese valor, consulta Tamaño del lienzo
 4. **Longitud**: `length` son fotogramas a 24fps (predeterminado **124** ≈ 5s), ajustado a **17k+5**
 5. **Texto en pantalla**: comillas dobles en inglés para cada cadena, literal (p. ej., `A red neon sign reading "营业中"`). Las líneas habladas usan `<d>`, no comillas
 6. **Marcadores de audio de R2V**: en el análisis de retención, marca cada referencia de audio como `fully_copy`, `partially_copy`, `reference` o `weak_reference` según cómo se use el clip (consulta la tabla de la guía oficial de referencias)
@@ -47,6 +48,22 @@ Escribe los prompts en **inglés**. Los diálogos y las letras de canciones van
 
 Las etiquetas y controles específicos de cada flujo de trabajo están en la página de cada flujo de trabajo.
 
+## Tamaño del lienzo
+
+Fija el lienzo antes de ajustar el prompt. Cambia `megapixels` en el nodo `Resolution Selector`; las plantillas abren por debajo del lienzo nativo:
+
+| Plantilla | Relación | Tamaño a 0.4 MP |
+| --- | --- | --- |
+| T2V | 16:9 | 864×480 |
+| I2V | 1:1 | 640×640 |
+| R2V | 16:9 | 864×480 |
+
+En 16:9 el lienzo nativo es **1344×768**, que la nota Size Settings Reference de la plantilla lista con **0.98** MP. El selector no lo limita por arriba, así que un lienzo mayor cuesta tiempo de render sin añadir detalle.
+
+El lienzo cambia el latente de vídeo, el latente de audio solo sigue la duración del clip, y ambos se generan en la misma pasada. Por eso el audio queda junto a un latente de vídeo con otra forma en cada tamaño: el mismo prompt y la misma semilla dan diálogos distintos y siguen el prompt en distinto grado. En R2V, un lienzo mayor además reduce menos las imágenes de referencia, así que la identidad se mantiene mejor.
+
+Para conservar una toma que ya funciona, amplía el clip terminado en lugar de volver a renderizarlo en un lienzo nuevo. Consulta [Escalado de vídeo](/es/tutorials/utility/video-upscale).
+
 ## Tomas y cortes
 
 Numera las tomas `[Shot 1]`, `[Shot 2]`, … `[Shot 1]` empieza en 0.00s. Las tomas posteriores comienzan con un momento de corte:
diff --git a/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
index 150c2ce703..5f32b8f5a6 100644
--- a/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
+++ b/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
@@ -2,12 +2,13 @@
 title: "ComfyUI MiniMax H3 プロンプトガイド: 公式ガイド、ヒント、プロンプト埋め込み"
 description: "ComfyUI でより良い MiniMax H3 プロンプトを書く方法: MiniMax の公式プロンプト作成ガイド、一般的なプロンプトのコツ、セリフと話者のルール、スタイル埋め込み。"
 sidebarTitle: "プロンプトガイド"
-translationSourceHash: 0f1ded35
+translationSourceHash: cdaa2314
 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx
 translationBlockHashes:
   "_intro": da22cc9d
   "Official guides": 3d762c67
-  "General tips": fab4ad35
+  "General tips": 447ee59b
+  "Canvas size": 8c15c46c
   "Shots and cuts": 8184dae6
   "Dialogue and speakers": 4985438e
   "Prompt embeddings": 9c899169
@@ -39,7 +40,7 @@ non_diegetic_music: ...
 
 1. **シーンを先に**: まず全体のシーンを書き、次に `integrated_multimodal_description` の `[Shot 1]` 内に時間指定付きのショットを書く
 2. **1 つのブロック**: ショット、カメラ、オーディオをまとめて記述する
-3. **解像度**: 16:9 でネイティブ **1344×768**、**32** の倍数
+3. **解像度**: 16:9 でネイティブ **1344×768**、**32** の倍数。テンプレートはそれより小さい値で開きます。キャンバスサイズを参照
 4. **長さ**: `length` は 24fps のフレーム数（既定 **124** ≈ 5s）で、**17k+5** にスナップされる
 5. **画面文字**: 各文字列を英語の二重引用符で囲み、原文のまま記述する（例: `A red neon sign reading "营业中"`）。話される台詞には `<d>` を使用し、引用符は使わない
 6. **R2V オーディオマーカー**: 保持分析では、クリップの使われ方に合わせて各オーディオ参照を `fully_copy`、`partially_copy`、`reference`、`weak_reference` のいずれかでマークする（公式の ref ガイドの表を参照）
@@ -49,6 +50,22 @@ non_diegetic_music: ...
 
 ワークフロー固有のタグとコントロールは各ワークフローページにあります。
 
+## キャンバスサイズ
+
+プロンプトを詰める前にキャンバスを決めてください。`Resolution Selector` ノードの `megapixels` を変更します。テンプレートはどれもネイティブのキャンバスより小さい値で開きます:
+
+| テンプレート | アスペクト比 | 0.4 MP のサイズ |
+| --- | --- | --- |
+| T2V | 16:9 | 864×480 |
+| I2V | 1:1 | 640×640 |
+| R2V | 16:9 | 864×480 |
+
+16:9 のネイティブキャンバスは **1344×768** で、テンプレートの Size Settings Reference ノートには **0.98** MP として載っています。セレクターはその上限で切り詰めないため、大きなキャンバスはディテールを増やさずレンダー時間だけを増やします。
+
+キャンバスを変えると変わるのは映像の latent で、音声の latent はクリップの長さにしか従いません。両ストリームは同じパスで生成されるため、音声はサイズごとに形の違う映像 latent の隣に並びます。同じプロンプトとシードでも台詞が変わり、プロンプトへの追従の度合いも変わります。R2V ではキャンバスを大きくすると参照画像の縮小も緩むため、同一性が保たれやすくなります。
+
+仕上がったテイクは再レンダーせず、クリップをアップスケールしてください。[動画のアップスケール](/ja/tutorials/utility/video-upscale)を参照。
+
 ## ショットとカット
 
 `[Shot 1]` は 0.00s から; 以降はカット時刻で開始:
diff --git a/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
index 9813eac4af..15423caf8d 100644
--- a/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
+++ b/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
@@ -2,12 +2,13 @@
 title: "ComfyUI MiniMax H3 프롬프트 가이드: 공식 가이드, 팁, 프롬프트 임베딩"
 description: "ComfyUI에서 더 나은 MiniMax H3 프롬프트 작성하기: MiniMax 공식 프롬프트 작성 가이드, 일반 프롬프트 팁, 대사와 화자 규칙, 스타일 임베딩."
 sidebarTitle: "프롬프트 가이드"
-translationSourceHash: 0f1ded35
+translationSourceHash: cdaa2314
 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx
 translationBlockHashes:
   "_intro": da22cc9d
   "Official guides": 3d762c67
-  "General tips": fab4ad35
+  "General tips": 447ee59b
+  "Canvas size": 8c15c46c
   "Shots and cuts": 8184dae6
   "Dialogue and speakers": 4985438e
   "Prompt embeddings": 9c899169
@@ -39,7 +40,7 @@ non_diegetic_music: ...
 
 1. **장면 우선**: 전체 장면을 먼저 쓰고, 그다음 `integrated_multimodal_description`의 `[Shot 1]` 안에 시간별 샷을 씁니다
 2. **하나의 블록**: 샷, 카메라, 오디오를 함께 씁니다
-3. **해상도**: 16:9에서 네이티브 **1344×768**; **32**의 배수
+3. **해상도**: 16:9에서 네이티브 **1344×768**; **32**의 배수. 템플릿은 그보다 작은 값으로 열립니다. 캔버스 크기를 참고하세요
 4. **길이**: `length`는 24fps 기준 프레임 수(기본 **124** ≈ 5s)이며 **17k+5**에 맞춰집니다
 5. **화면 텍스트**: 문자열마다 영어 큰따옴표로, 원문 그대로(예: `A red neon sign reading "营业中"`). 대사는 큰따옴표가 아니라 `<d>`를 씁니다
 6. **R2V 오디오 마커**: 유지 분석에서 각 오디오 참조를 `fully_copy`, `partially_copy`, `reference`, `weak_reference` 중 하나로 표시해 클립이 사용되는 방식과 일치시키세요(공식 ref 가이드 표 참고)
@@ -49,6 +50,22 @@ non_diegetic_music: ...
 
 워크플로별 태그와 컨트롤은 각 워크플로 페이지에 있습니다.
 
+## 캔버스 크기
+
+프롬프트를 다듬기 전에 캔버스를 먼저 정하세요. `Resolution Selector` 노드의 `megapixels`를 바꿉니다. 템플릿은 모두 네이티브 캔버스보다 작은 값으로 열립니다:
+
+| 템플릿 | 종횡비 | 0.4 MP 크기 |
+| --- | --- | --- |
+| T2V | 16:9 | 864×480 |
+| I2V | 1:1 | 640×640 |
+| R2V | 16:9 | 864×480 |
+
+16:9의 네이티브 캔버스는 **1344×768**이며, 템플릿의 Size Settings Reference 노트에 **0.98** MP로 적혀 있습니다. 선택기는 그 위에서 값을 제한하지 않으므로, 더 큰 캔버스는 디테일 없이 렌더 시간만 늘립니다.
+
+캔버스를 바꾸면 바뀌는 것은 비디오 latent이고, 오디오 latent는 클립 길이만 따릅니다. 두 스트림은 같은 패스에서 생성되므로, 오디오는 크기마다 모양이 다른 비디오 latent 옆에 놓입니다. 같은 프롬프트와 시드라도 대사가 달라지고 프롬프트를 따르는 정도도 달라집니다. R2V에서는 캔버스가 커지면 참조 이미지가 덜 축소되어 동일성이 더 잘 유지됩니다.
+
+이미 잘 나온 테이크는 다시 렌더하지 말고 클립을 업스케일하세요. [비디오 업스케일](/ko/tutorials/utility/video-upscale)을 참고하세요.
+
 ## 샷과 컷
 
 `[Shot 1]` 0.00s부터; 이후는 컷 시각으로 시작:
diff --git a/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
index 2c4b59cd6b..0c56a71716 100644
--- a/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
+++ b/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
@@ -29,7 +29,7 @@ Write prompts in **English**. Dialogue/lyrics in `<d>` tags; on-screen text verb
 
 1. **Scene first**: Overall scene, then timed shots inside `[Shot 1]` of `integrated_multimodal_description`
 2. **One block**: Shots, camera, and audio together
-3. **Resolution**: Native **1344×768** at 16:9; multiples of **32**
+3. **Resolution**: Native **1344×768** at 16:9; multiples of **32**. The templates open below that, see Canvas size
 4. **Length**: `length` is frames at 24fps (default **124** ≈ 5s), snapped to **17k+5**
 5. **On-screen text**: English double quotes per string, verbatim (e.g. `A red neon sign reading "营业中"`). Spoken lines use `<d>`, not quotes
 6. **R2V audio markers**: In retention analysis, mark each audio ref as `fully_copy`, `partially_copy`, `reference`, or `weak_reference` to match how the clip is used (see official ref guide table)
@@ -39,6 +39,22 @@ Write prompts in **English**. Dialogue/lyrics in `<d>` tags; on-screen text verb
 
 Workflow-specific tags and controls live on each workflow page.
 
+## Canvas size
+
+Set the canvas before you tune the prompt. Change `megapixels` on the `Resolution Selector` node; every template opens below the native canvas:
+
+| Template | Aspect | Size at 0.4 MP |
+| --- | --- | --- |
+| T2V | 16:9 | 864×480 |
+| I2V | 1:1 | 640×640 |
+| R2V | 16:9 | 864×480 |
+
+For 16:9 the native canvas is **1344×768**, which the template's Size Settings Reference note lists at **0.98** MP. The selector does not clamp above it, so a larger canvas costs render time without adding detail.
+
+Canvas size changes the video latent, the audio latent follows clip length only, and both are generated in the same pass. The audio therefore sits beside a differently shaped video latent at every size, so the same prompt and seed give different dialogue and follow the prompt to a different degree. In R2V, a larger canvas also shrinks the reference images less, so identity holds better.
+
+To keep a take that already works, upscale the finished clip instead of re-rendering it on a new canvas. See [Video upscale](/tutorials/utility/video-upscale).
+
 ## Shots and cuts
 
 Number shots `[Shot 1]`, `[Shot 2]`, … `[Shot 1]` starts at 0.00s. Later shots open with a cut time:
diff --git a/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
index 7dca33331a..81631c3c84 100644
--- a/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
+++ b/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
@@ -2,12 +2,13 @@
 title: "ComfyUI MiniMax H3 提示词指南：官方指南、技巧与提示词嵌入"
 description: "在 ComfyUI 中写出更好的 MiniMax H3 提示词：MiniMax 官方提示词编写指南、通用提示词技巧、对白与说话人规则，以及风格嵌入。"
 sidebarTitle: "提示词指南"
-translationSourceHash: 0f1ded35
+translationSourceHash: cdaa2314
 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx
 translationBlockHashes:
   "_intro": da22cc9d
   "Official guides": 3d762c67
-  "General tips": fab4ad35
+  "General tips": 447ee59b
+  "Canvas size": 8c15c46c
   "Shots and cuts": 8184dae6
   "Dialogue and speakers": 4985438e
   "Prompt embeddings": 9c899169
@@ -39,7 +40,7 @@ non_diegetic_music: ...
 
 1. **先写场景**：先写整体场景，再在 `integrated_multimodal_description` 的 `[Shot 1]` 内按时间顺序写分镜
 2. **一个整体**：分镜、相机与音频放在一起写
-3. **分辨率**：16:9 原生 **1344×768**；取 **32** 的倍数
+3. **分辨率**：16:9 原生 **1344×768**；取 **32** 的倍数。模板默认的画布低于这个值，见画布尺寸
 4. **长度**：`length` 是以 24fps 计的帧数（默认 **124** ≈ 5s），会对齐到 **17k+5**
 5. **屏幕文字**：每个字符串用英文双引号包裹，逐字照抄（如 `A red neon sign reading "营业中"`）。台词使用 `<d>`，不要用引号
 6. **R2V 音频标记**：在保留分析中，按片段的实际用法把每个音频参考标记为 `fully_copy`、`partially_copy`、`reference` 或 `weak_reference`（见官方参考指南表格）
@@ -49,6 +50,22 @@ non_diegetic_music: ...
 
 各工作流专属的标签与控制项见对应工作流页面。
 
+## 画布尺寸
+
+请在调提示词之前先定好画布。修改 `Resolution Selector` 节点上的 `megapixels`，模板的默认画布都低于原生画布：
+
+| 模板 | 宽高比 | 0.4 MP 下的尺寸 |
+| --- | --- | --- |
+| T2V | 16:9 | 864×480 |
+| I2V | 1:1 | 640×640 |
+| R2V | 16:9 | 864×480 |
+
+16:9 的原生画布是 **1344×768**，模板里的 Size Settings Reference 注释把它列在 **0.98** MP。选择器不会把尺寸限制在这个上限之内，更大的画布只多花渲染时间，不增加细节。
+
+改画布改的是视频 latent，音频 latent 只随片段长度变化，两者在同一次采样中生成。因此音频在每个尺寸下都挨着形状不同的视频 latent，同一提示词与种子会得到不同的对白，对提示词的遵循程度也会变。R2V 中画布变大还会让参考图少缩一些，主体一致性更好。
+
+已经满意的一版不要重渲：把成片放大，而不是换一块新画布重跑。见[视频放大](/zh/tutorials/utility/video-upscale)。
+
 ## 镜头与切换
 
 `[Shot 1]` 从 0.00s 起；后续镜头以切换时间开头：

From 255775ce9d3bb2a3f39d09152ccf49104d49e7b8 Mon Sep 17 00:00:00 2001
From: lin-bot23 <linzhixiong23@gmail.com>
Date: Sun, 1 Nov 2026 18:03:56 +0800
Subject: [PATCH 09/13] docs(minimax): document H3 context windows for long
 clips

---
 .../video/minimax/minimax-h3-native.mdx       | 24 +++++++++++++++++--
 .../video/minimax/minimax-h3-native.mdx       | 24 +++++++++++++++++--
 .../video/minimax/minimax-h3-native.mdx       | 24 +++++++++++++++++--
 tutorials/video/minimax/minimax-h3-native.mdx | 20 ++++++++++++++++
 .../video/minimax/minimax-h3-native.mdx       | 24 +++++++++++++++++--
 5 files changed, 108 insertions(+), 8 deletions(-)

diff --git a/es/tutorials/video/minimax/minimax-h3-native.mdx b/es/tutorials/video/minimax/minimax-h3-native.mdx
index 5f1c134b4a..1c4348a169 100644
--- a/es/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/es/tutorials/video/minimax/minimax-h3-native.mdx
@@ -2,14 +2,14 @@
 title: "Flujos de trabajo de MiniMax H3 en ComfyUI: texto a video, imagen a video y referencia a video"
 description: "Ejemplos paso a paso de flujos nativos de MiniMax H3 en ComfyUI, con anclaje de guías y máscaras de ruido latente para un control más fino del video."
 sidebarTitle: "Flujos nativos"
-translationSourceHash: f8176106
+translationSourceHash: 2a225f15
 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx
 translationBlockHashes:
   "_intro": dde651e5
   "MiniMax H3 Text to Video (T2V)": 42ec8cfb
   "MiniMax H3 Image to Video (I2V)": 2a06e3fa
   "MiniMax H3 Reference to Video (R2V)": 961829df
-  "Advanced workflows with the native nodes": 85460bff
+  "Advanced workflows with the native nodes": bbaed645
 ---
 Esta página cubre las tres plantillas base de MiniMax H3 y las técnicas avanzadas con nodos nativos. Para conocer las capacidades y el licenciamiento, consulta la [descripción general de MiniMax H3](/es/tutorials/video/minimax/minimax-h3).
 
@@ -258,6 +258,26 @@ Las [máscaras de ruido por token](https://github.com/Comfy-Org/ComfyUI/pull/153
 
 Conecta una máscara a `denoise_mask` del sampler: **0** = mantener, **1** = regenerar. Las máscaras de video se ajustan a la cuadrícula de parches latentes de 2×2; las de audio, a frames latentes completos. Úsalas para relleno inteligente, eliminación o extensión mientras mantienes el resto estable.
 
+### Clips largos con ventanas de contexto (#17425)
+
+[`MiniMaxH3ContextWindows`](https://github.com/Comfy-Org/ComfyUI/pull/17425) muestrea un clip largo en ventanas superpuestas, así que cada paso mantiene una ventana en memoria en lugar del latente completo. Conéctalo entre `UNETLoader` y `BasicGuider` (model a la entrada, model a la salida), el mismo sitio que cualquier otro parche de modelo.
+
+| Control | Significado | Valor por defecto |
+| --- | --- | --- |
+| `context_length` | Tamaño de la ventana en frames, **17n+5** | **124**, la longitud del clip por defecto |
+| `context_overlap` | Frames que comparten ventanas contiguas, **17n+5** | **39** |
+| `fuse_method` | Cómo se combinan las ventanas superpuestas | `pyramid` |
+| `freenoise` | Mezcla el ruido entre ventanas para unirlas | Activado |
+
+Las ventanas empiezan en los bloques de 17 frames del modelo, de modo que video y audio siguen alineados al cruzar un límite. El nodo mantiene `context_overlap` por debajo de la longitud de la ventana. Con el valor por defecto **124** el clip de 5 segundos cabe en una sola ventana, así que el nodo solo cambia el muestreo cuando el clip supera la ventana: baja `context_length` para que quepa un clip más largo en la misma tarjeta, súbelo para dar más contexto a cada paso a cambio de más memoria.
+
+El muestreo por ventanas trae dos límites:
+
+- **El prompt se ejecuta completo dentro de cada ventana.** Los tiempos de plano se leen por ventana, no sobre el clip entero, así que un corte con marca de tiempo vuelve a dispararse en cada una. Describe una escena continua y deja fuera los cortes con tiempo.
+- **Las ventanas se fusionan, no se empalman.** El contenido que debe mantenerse idéntico al cruzar un límite (texto en pantalla, detalle del decorado, una cara) puede derivar. Mantén estables la cámara y el sujeto.
+
+Por eso las ventanas de contexto encajan en una escena sin cortes: un push lento, un paneo o un segundo pase de escalado sobre un clip terminado. Para un clip que cambia de plano, genera los planos por separado y únelos.
+
 ### Generar primero la banda sonora
 
 El audio y el video se desruidizan juntos, así que una programación corta o un LoRA turbo suavizan la pista junto con la imagen. Dos pasadas los separan: fija el audio en un lienzo pequeño y luego condiciona el video con él.
diff --git a/ja/tutorials/video/minimax/minimax-h3-native.mdx b/ja/tutorials/video/minimax/minimax-h3-native.mdx
index 8feae41a0b..b6bbc7fd2a 100644
--- a/ja/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/ja/tutorials/video/minimax/minimax-h3-native.mdx
@@ -2,14 +2,14 @@
 title: "ComfyUI MiniMax H3 テキストから動画・画像から動画・参照から動画のワークフロー"
 description: "ComfyUI における MiniMax H3 ネイティブワークフローの手順例。ガイドのアンカー固定や latent ノイズマスクで生成動画をより細かく制御できます。"
 sidebarTitle: "ネイティブワークフロー"
-translationSourceHash: f8176106
+translationSourceHash: 2a225f15
 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx
 translationBlockHashes:
   "_intro": dde651e5
   "MiniMax H3 Text to Video (T2V)": 42ec8cfb
   "MiniMax H3 Image to Video (I2V)": 2a06e3fa
   "MiniMax H3 Reference to Video (R2V)": 961829df
-  "Advanced workflows with the native nodes": 85460bff
+  "Advanced workflows with the native nodes": bbaed645
 ---
 
 このページは 3 つの基本 MiniMax H3 テンプレートとネイティブノードの高度なテクニックを扱います。機能とライセンスは [MiniMax H3 概要](/ja/tutorials/video/minimax/minimax-h3) を参照してください。
@@ -259,6 +259,26 @@ H3 ノードの `positive` と latent を `MiniMaxH3AddGuide` に接続:
 
 サンプラー `denoise_mask` にマスク: **0** 保持、**1** 再生成。動画は 2×2 latent パッチ、音声は latent フレーム単位。インペイント、削除、拡張に使用。
 
+### コンテキストウィンドウで長いクリップを生成 (#17425)
+
+[`MiniMaxH3ContextWindows`](https://github.com/Comfy-Org/ComfyUI/pull/17425) は長いクリップを重なり合うウィンドウに分割してサンプリングし、各ステップで全体ではなく 1 ウィンドウ分だけをメモリに載せます。`UNETLoader` と `BasicGuider` の間（model 入力、model 出力）に接続します。他の model パッチと同じ位置です。
+
+| 項目 | 意味 | デフォルト |
+| --- | --- | --- |
+| `context_length` | ウィンドウ長（フレーム）、**17n+5** | **124**（デフォルトのクリップ長） |
+| `context_overlap` | 隣接ウィンドウが共有するフレーム数、**17n+5** | **39** |
+| `fuse_method` | 重なったウィンドウの融合方法 | `pyramid` |
+| `freenoise` | ウィンドウ間でノイズをシャッフルしてつなぐ | 有効 |
+
+ウィンドウはモデルの 17 フレーム塊の境界から始まるため、動画と音声は境界をまたいでも揃います。ノードは `context_overlap` をウィンドウ長未満に抑えます。デフォルトの **124** では 5 秒のクリップ全体が 1 ウィンドウに入るので、クリップがウィンドウを超えてから初めて効果が出ます。`context_length` を下げると同じ GPU で長いクリップが入り、上げると各ステップの前後文脈が増える代わりにメモリを多く使います。
+
+ウィンドウ分割には 2 つの制約があります:
+
+- **各ウィンドウでプロンプト全体が実行されます。** ショットの時刻はクリップ全体ではなくウィンドウ基準で読まれるため、時刻指定のカットは各ウィンドウで再度発生します。1 つの連続したシーンとして書き、時刻指定のカットは入れません。
+- **ウィンドウは融合されるのであって連結ではありません。** 境界をまたいで厳密に保つ必要がある内容（画面内の文字、セットの細部、顔）はドリフトし得ます。カメラと被写体は安定させます。
+
+そのためコンテキストウィンドウは、カットのない連続したシーン（ゆっくりしたプッシュイン、パン、完成クリップの 2 パスアップスケール）に向きます。ショットを切り替えるクリップは、ショットごとに生成してからつなぎます。
+
 ### 音声を先に作る
 
 音声と動画は同じデノイズで生成されるため、短いスケジュールや turbo LoRA は映像と一緒に音声も甘くします。2 パスに分ければ切り離せます。まず小さいキャンバスで音声を固め、その音声を条件に映像を作ります。
diff --git a/ko/tutorials/video/minimax/minimax-h3-native.mdx b/ko/tutorials/video/minimax/minimax-h3-native.mdx
index ab26e4077f..a72c906344 100644
--- a/ko/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/ko/tutorials/video/minimax/minimax-h3-native.mdx
@@ -2,14 +2,14 @@
 title: "ComfyUI MiniMax H3 텍스트·이미지·레퍼런스 기반 비디오 워크플로"
 description: "ComfyUI의 MiniMax H3 네이티브 워크플로 단계별 예제입니다. 가이드 고정과 잠재 노이즈 마스크로 생성 비디오를 더 세밀하게 제어할 수 있습니다."
 sidebarTitle: "네이티브 워크플로"
-translationSourceHash: f8176106
+translationSourceHash: 2a225f15
 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx
 translationBlockHashes:
   "_intro": dde651e5
   "MiniMax H3 Text to Video (T2V)": 42ec8cfb
   "MiniMax H3 Image to Video (I2V)": 2a06e3fa
   "MiniMax H3 Reference to Video (R2V)": 961829df
-  "Advanced workflows with the native nodes": 85460bff
+  "Advanced workflows with the native nodes": bbaed645
 ---
 
 이 페이지는 세 가지 기본 MiniMax H3 템플릿과 네이티브 노드 고급 기법을 다룹니다. 기능과 라이선스는 [MiniMax H3 개요](/ko/tutorials/video/minimax/minimax-h3)를 참조하세요.
@@ -259,6 +259,26 @@ H3 노드 `positive`·latent를 `MiniMaxH3AddGuide`에 연결:
 
 샘플러 `denoise_mask`: **0** 유지, **1** 재생성. 비디오 2×2 latent 패치, 오디오 latent 프레임 단위. 인페인팅, 제거, 확장.
 
+### 컨텍스트 윈도우로 긴 클립 생성 (#17425)
+
+[`MiniMaxH3ContextWindows`](https://github.com/Comfy-Org/ComfyUI/pull/17425)는 긴 클립을 겹치는 윈도우로 나눠 샘플링합니다. 각 단계에서 전체 latent 대신 윈도우 하나만 메모리에 올립니다. `UNETLoader`와 `BasicGuider` 사이(model 입력, model 출력)에 연결합니다. 다른 model 패치와 같은 자리입니다.
+
+| 항목 | 의미 | 기본값 |
+| --- | --- | --- |
+| `context_length` | 윈도우 길이(프레임), **17n+5** | **124**, 기본 클립 길이 |
+| `context_overlap` | 인접 윈도우가 공유하는 프레임 수, **17n+5** | **39** |
+| `fuse_method` | 겹친 윈도우를 합치는 방식 | `pyramid` |
+| `freenoise` | 윈도우 사이 노이즈를 섞어 이어 붙임 | 켜짐 |
+
+윈도우는 모델의 17프레임 덩어리 경계에서 시작하므로 영상과 오디오가 경계를 넘어도 어긋나지 않습니다. 노드는 `context_overlap`을 윈도우 길이 아래로 제한합니다. 기본값 **124**에서는 5초 클립 전체가 윈도우 하나에 들어가므로, 클립이 윈도우보다 길어져야 효과가 생깁니다. `context_length`를 낮추면 같은 GPU에서 더 긴 클립이 들어가고, 높이면 각 단계가 더 많은 앞뒤 맥락을 보는 대신 메모리를 더 씁니다.
+
+윈도우 분할에는 제약이 두 가지 있습니다:
+
+- **모든 윈도우에서 프롬프트 전체가 실행됩니다.** 샷 시각은 클립 전체가 아니라 윈도우 기준으로 읽히므로, 시각을 지정한 컷은 윈도우마다 다시 발생합니다. 하나의 연속된 장면으로 쓰고 시각 지정 컷은 넣지 않습니다.
+- **윈도우는 이어 붙이는 게 아니라 합쳐집니다.** 경계를 넘어 정확히 유지해야 하는 내용(화면 속 글자, 세트 디테일, 얼굴)은 밀릴 수 있습니다. 카메라와 피사체는 안정적으로 유지합니다.
+
+그래서 컨텍스트 윈도우는 컷이 없는 연속 장면에 맞습니다. 느린 푸시인, 팬, 완성된 클립의 2패스 업스케일 같은 경우입니다. 샷을 전환하는 클립은 샷별로 생성한 뒤 이어 붙입니다.
+
 ### 사운드트랙을 먼저 만들기
 
 오디오와 비디오는 같은 디노이즈에서 함께 생성되므로, 짧은 스케줄이나 turbo LoRA는 영상과 함께 오디오도 무디게 만듭니다. 두 번의 패스로 분리할 수 있습니다. 작은 캔버스에서 오디오를 먼저 확정하고, 그 오디오를 조건으로 영상을 만듭니다.
diff --git a/tutorials/video/minimax/minimax-h3-native.mdx b/tutorials/video/minimax/minimax-h3-native.mdx
index 1af9f4e235..da9c0202ea 100644
--- a/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/tutorials/video/minimax/minimax-h3-native.mdx
@@ -251,6 +251,26 @@ See the [Multiframe Reference](/tutorials/video/minimax/minimax-h3-multiframe) t
 
 Connect a mask to the sampler `denoise_mask`: **0** = keep, **1** = regenerate. Video masks snap to the 2×2 latent patch grid; audio masks to whole latent frames. Use for inpainting, removal, or extension while holding the rest stable.
 
+### Long clips with context windows (#17425)
+
+[`MiniMaxH3ContextWindows`](https://github.com/Comfy-Org/ComfyUI/pull/17425) samples a long clip in overlapping windows, so each step holds one window in memory instead of the whole latent. Wire it between `UNETLoader` and `BasicGuider` (model in, model out), the same place as any other model patch.
+
+| Control | Meaning | Default |
+| --- | --- | --- |
+| `context_length` | Window size in frames, **17n+5** | **124**, the default clip length |
+| `context_overlap` | Frames shared by adjacent windows, **17n+5** | **39** |
+| `fuse_method` | How overlapping windows are combined | `pyramid` |
+| `freenoise` | Shuffles noise across windows to blend them | On |
+
+Windows start on the model's 17-frame chunks, so video and audio stay in step across a boundary. The node keeps `context_overlap` below the window length. At the default **124** the whole 5-second clip is a single window, so the node only changes sampling once the clip runs past the window: lower `context_length` to fit a longer clip on the same card, raise it to give each step more surrounding context at a higher memory cost.
+
+Two limits come with windowed sampling:
+
+- **The prompt runs in full inside every window.** Shot timings are read per window, not across the clip, so a timed cut fires again in each one. Describe one continuous scene and leave timed cuts out.
+- **Windows are fused, not joined.** Content that has to hold exactly across a boundary (on-screen text, set detail, a face) can drift. Keep the camera and the subject steady.
+
+That makes context windows a fit for a scene that runs without a cut: a slow push, a pan, or a second-pass upscale of a finished clip. For a clip that cuts between shots, render the shots separately and join them.
+
 ### Generating the soundtrack first
 
 Audio and video are denoised together, so a short schedule or a turbo LoRA softens the track along with the picture. Two passes separate them: settle the audio on a small canvas, then condition the video on it.
diff --git a/zh/tutorials/video/minimax/minimax-h3-native.mdx b/zh/tutorials/video/minimax/minimax-h3-native.mdx
index 3ca6b014ff..1a01b633ba 100644
--- a/zh/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/zh/tutorials/video/minimax/minimax-h3-native.mdx
@@ -2,14 +2,14 @@
 title: "ComfyUI MiniMax H3 文生视频、图生视频与参考生视频工作流"
 description: "ComfyUI 中 MiniMax H3 原生工作流的逐步示例，含帧锚定与潜在噪声蒙版技巧，帮助你对生成视频做更精细的控制。"
 sidebarTitle: "原生工作流"
-translationSourceHash: f8176106
+translationSourceHash: 2a225f15
 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx
 translationBlockHashes:
   "_intro": dde651e5
   "MiniMax H3 Text to Video (T2V)": 42ec8cfb
   "MiniMax H3 Image to Video (I2V)": 2a06e3fa
   "MiniMax H3 Reference to Video (R2V)": 961829df
-  "Advanced workflows with the native nodes": 85460bff
+  "Advanced workflows with the native nodes": bbaed645
 ---
 本页涵盖三个基础 MiniMax H3 模板及原生节点高级技巧。能力与授权见 [MiniMax H3 概览](/zh/tutorials/video/minimax/minimax-h3)。
 
@@ -258,6 +258,26 @@ ComfyUI/
 
 遮罩接采样器 `denoise_mask`：**0** 保留，**1** 重生成。视频对齐 2×2 latent 块，音频对齐整帧。用于修补、移除或扩展。
 
+### 用上下文窗口生成长片段 (#17425)
+
+[`MiniMaxH3ContextWindows`](https://github.com/Comfy-Org/ComfyUI/pull/17425) 把长片段切成互相重叠的窗口采样，每一步只把一个窗口放进显存，不必放下整个 latent。接在 `UNETLoader` 与 `BasicGuider` 之间（模型进、模型出），与其他模型 patch 的位置相同。
+
+| 控件 | 含义 | 默认值 |
+| --- | --- | --- |
+| `context_length` | 窗口长度（帧），**17n+5** | **124**，即默认片段长度 |
+| `context_overlap` | 相邻窗口共享的帧数，**17n+5** | **39** |
+| `fuse_method` | 重叠窗口的融合方式 | `pyramid` |
+| `freenoise` | 跨窗口打乱噪波以衔接 | 开启 |
+
+窗口从模型的 17 帧块边界开始，视频与音频跨边界仍对齐。节点会把 `context_overlap` 压在窗口长度以下。默认 **124** 时整段 5 秒片段就是一个窗口，片段超过窗口长度后节点才真正起作用：调小 `context_length` 让更长的片段塞进同一张卡，调大则每一步拿到更多上下文，显存占用随之上升。
+
+窗口化采样自带两个限制：
+
+- **每个窗口都会跑完整提示词。** 分镜时间按窗口读，不是按整段读，定时切镜会在每个窗口里再触发一次。写一个连续的场景，不要写定时切镜。
+- **窗口是融合，不是拼接。** 需要跨边界保持一致的内容（画面文字、场景细节、人脸）可能漂移。镜头和主体保持稳定。
+
+所以上下文窗口适合没有切镜的连续场景：缓慢推镜、平移，或对成片做二次放大。需要切镜的片段，分镜单独生成再拼接。
+
 ### 先生成音轨
 
 音频与视频在同一次去噪中生成，短调度或 turbo LoRA 会连同画面一起把音轨变软。分两遍就能拆开：先在小画布上把音频定下来，再让视频以它为条件。

From 231d5dadd8656f3beef7faf5bf432543c530a272 Mon Sep 17 00:00:00 2001
From: lin-bot23 <linzhixiong23@gmail.com>
Date: Sun, 1 Nov 2026 18:04:08 +0800
Subject: [PATCH 10/13] chore: restore generated router pages to upstream

---
 development/comfy-router/limitations.mdx      |  6 ++----
 .../models/runway/gen4-image/code.mdx         |  2 +-
 .../xai/grok-imagine-image-2-0/code.mdx       |  4 ++--
 .../xai/grok-imagine-image-pro/code.mdx       |  4 ++--
 .../xai/grok-imagine-image-quality/code.mdx   |  4 ++--
 .../models/xai/grok-imagine-image/code.mdx    |  4 ++--
 development/comfy-router/reference.mdx        |  6 +++---
 .../video/minimax/minimax-h3-native.mdx       | 20 ++++++++++---------
 .../video/minimax/minimax-h3-native.mdx       | 20 ++++++++++---------
 .../video/minimax/minimax-h3-native.mdx       | 20 ++++++++++---------
 router-openapi.yaml                           |  2 +-
 router-schemas/runway/gen4_image.json         |  2 +-
 .../xai/grok-imagine-image-2.0.json           |  2 +-
 .../xai/grok-imagine-image-pro.json           |  2 +-
 .../xai/grok-imagine-image-quality.json       |  2 +-
 router-schemas/xai/grok-imagine-image.json    |  2 +-
 tutorials/video/minimax/minimax-h3-native.mdx | 16 ++++++++-------
 .../video/minimax/minimax-h3-native.mdx       | 20 ++++++++++---------
 18 files changed, 73 insertions(+), 65 deletions(-)

diff --git a/development/comfy-router/limitations.mdx b/development/comfy-router/limitations.mdx
index c370edb495..9fd858db81 100644
--- a/development/comfy-router/limitations.mdx
+++ b/development/comfy-router/limitations.mdx
@@ -92,13 +92,11 @@ This list shrinks as estimate coverage grows. It is generated from Router's own
 {/* BEGIN GENERATED: models without a pre-flight estimate. Do not hand-edit; regenerate with -update-quote-coverage-docs. */}
 
 - **bfl**: `bfl/erase-v1`, `bfl/video-edit-v1`, `bfl/video-upscale-v1`, `bfl/vto-v1`
-- **fal**: `fal/crystal-upscaler`, `fal/crystal-video-upscaler`, `fal/fal-gpt-image-2`, `fal/fal-gpt-image-2.5-flare`, `fal/fal-gpt-image-2.5-sunburst`, `fal/fal-nano-banana-2`, `fal/fal-nano-banana-pro`, `fal/fal-seedance-2.0`, `fal/fal-seedance-2.5`, `fal/flux-3-max`, `fal/patina`
+- **fal**: `fal/crystal-upscaler`, `fal/crystal-video-upscaler`, `fal/fal-seedance-2.0`, `fal/fal-seedance-2.5`, `fal/patina`
 - **freepik**: `freepik/ai-image-upscaler-precision-v2`, `freepik/ai-skin-enhancer-creative`, `freepik/ai-skin-enhancer-faithful`, `freepik/ai-skin-enhancer-flexible`
 - **heygen**: `heygen/starfish`
 - **ideogram**: `ideogram/ideogram-v3`, `ideogram/ideogram-v4`, `ideogram/p-image-ideogram`
-- **kling**: `kling/kling-3.0-turbo`, `kling/kling-image-o1`, `kling/kling-v2-5-turbo`, `kling/kling-v2-6`, `kling/kling-v3`, `kling/kling-v3-omni`, `kling/kling-video-o1`, `kling/videos-avatar-image2video`, `kling/videos-lip-sync`, `kling/videos-video-extend`
-- **luma**: `luma/photon-1`, `luma/photon-flash-1`
-- **meshy**: `meshy/animations`, `meshy/meshy-6`, `meshy/meshy-7`, `meshy/meshy-7.1`, `meshy/remesh`, `meshy/rigging`
+- **kling**: `kling/kling-image-o1`, `kling/videos-avatar-image2video`, `kling/videos-lip-sync`, `kling/videos-video-extend`
 - **moonvalley**: `moonvalley/image-to-video`, `moonvalley/text-to-video`, `moonvalley/video-to-video`
 - **openai**: `openai/gpt-image-1.5`
 - **openrouter**: `openrouter/chat-completions`
diff --git a/development/comfy-router/models/runway/gen4-image/code.mdx b/development/comfy-router/models/runway/gen4-image/code.mdx
index 3a57a6de31..1d1875cf10 100644
--- a/development/comfy-router/models/runway/gen4-image/code.mdx
+++ b/development/comfy-router/models/runway/gen4-image/code.mdx
@@ -160,7 +160,7 @@ curl https://api.comfy.org/v2/models/runway/gen4_image/requests/$REQUEST_ID \
 </ParamField>
 
 <ParamField body="referenceImages[].uri" type="string" required>
-  A HTTPS URL or data URI containing an encoded image
+  A HTTPS URL, Runway upload URI (runway://) or data URI containing an encoded image
 </ParamField>
 
 Generated from the schema Router serves at `GET /v2/models/runway/gen4_image/openapi.json`, the same document it validates a call against before the request reaches the provider.
diff --git a/development/comfy-router/models/xai/grok-imagine-image-2-0/code.mdx b/development/comfy-router/models/xai/grok-imagine-image-2-0/code.mdx
index 24ab6bb30f..4ae1e2dc33 100644
--- a/development/comfy-router/models/xai/grok-imagine-image-2-0/code.mdx
+++ b/development/comfy-router/models/xai/grok-imagine-image-2-0/code.mdx
@@ -164,9 +164,9 @@ curl https://api.comfy.org/v2/models/xai/grok-imagine-image-2.0/requests/$REQUES
 </ParamField>
 
 <ParamField body="quality" type="string">
-  Quality of the output image. For grok-imagine-image-2.0 this selects the price tier (low/medium; medium is the default); other models currently ignore it.
+  Quality of the output image. For grok-imagine-image-2.0 this selects the price tier: low, medium, or auto (xAI's default when omitted; auto serves low for generation and medium for edits). Other models ignore it.
 
-  Possible values: `low`, `medium`, `high`
+  Possible values: `low`, `medium`, `high`, `auto`
 </ParamField>
 
 <ParamField body="resolution" type="string" default="&quot;1k&quot;">
diff --git a/development/comfy-router/models/xai/grok-imagine-image-pro/code.mdx b/development/comfy-router/models/xai/grok-imagine-image-pro/code.mdx
index 23d3c39d17..4f1c1e678d 100644
--- a/development/comfy-router/models/xai/grok-imagine-image-pro/code.mdx
+++ b/development/comfy-router/models/xai/grok-imagine-image-pro/code.mdx
@@ -164,9 +164,9 @@ curl https://api.comfy.org/v2/models/xai/grok-imagine-image-pro/requests/$REQUES
 </ParamField>
 
 <ParamField body="quality" type="string">
-  Quality of the output image. For grok-imagine-image-2.0 this selects the price tier (low/medium; medium is the default); other models currently ignore it.
+  Quality of the output image. For grok-imagine-image-2.0 this selects the price tier: low, medium, or auto (xAI's default when omitted; auto serves low for generation and medium for edits). Other models ignore it.
 
-  Possible values: `low`, `medium`, `high`
+  Possible values: `low`, `medium`, `high`, `auto`
 </ParamField>
 
 <ParamField body="resolution" type="string" default="&quot;1k&quot;">
diff --git a/development/comfy-router/models/xai/grok-imagine-image-quality/code.mdx b/development/comfy-router/models/xai/grok-imagine-image-quality/code.mdx
index 75ce14f7ce..2090987cd8 100644
--- a/development/comfy-router/models/xai/grok-imagine-image-quality/code.mdx
+++ b/development/comfy-router/models/xai/grok-imagine-image-quality/code.mdx
@@ -164,9 +164,9 @@ curl https://api.comfy.org/v2/models/xai/grok-imagine-image-quality/requests/$RE
 </ParamField>
 
 <ParamField body="quality" type="string">
-  Quality of the output image. For grok-imagine-image-2.0 this selects the price tier (low/medium; medium is the default); other models currently ignore it.
+  Quality of the output image. For grok-imagine-image-2.0 this selects the price tier: low, medium, or auto (xAI's default when omitted; auto serves low for generation and medium for edits). Other models ignore it.
 
-  Possible values: `low`, `medium`, `high`
+  Possible values: `low`, `medium`, `high`, `auto`
 </ParamField>
 
 <ParamField body="resolution" type="string" default="&quot;1k&quot;">
diff --git a/development/comfy-router/models/xai/grok-imagine-image/code.mdx b/development/comfy-router/models/xai/grok-imagine-image/code.mdx
index 4452f973dd..1cfa927dbb 100644
--- a/development/comfy-router/models/xai/grok-imagine-image/code.mdx
+++ b/development/comfy-router/models/xai/grok-imagine-image/code.mdx
@@ -164,9 +164,9 @@ curl https://api.comfy.org/v2/models/xai/grok-imagine-image/requests/$REQUEST_ID
 </ParamField>
 
 <ParamField body="quality" type="string">
-  Quality of the output image. For grok-imagine-image-2.0 this selects the price tier (low/medium; medium is the default); other models currently ignore it.
+  Quality of the output image. For grok-imagine-image-2.0 this selects the price tier: low, medium, or auto (xAI's default when omitted; auto serves low for generation and medium for edits). Other models ignore it.
 
-  Possible values: `low`, `medium`, `high`
+  Possible values: `low`, `medium`, `high`, `auto`
 </ParamField>
 
 <ParamField body="resolution" type="string" default="&quot;1k&quot;">
diff --git a/development/comfy-router/reference.mdx b/development/comfy-router/reference.mdx
index ea238ba01b..34f07809ac 100644
--- a/development/comfy-router/reference.mdx
+++ b/development/comfy-router/reference.mdx
@@ -975,7 +975,7 @@ Where the estimate fails closed:
 Limits of the estimate. A `200` figure is not a promise that the run is accepted, or that it is charged exactly that:
 
 - `/estimate` takes no `strict_mode` or `reject_unknown_fields`, so it checks the model's native contract only: a body a `strict_mode=true` run accepts in the alternate provider's own schema can get a `422` here, and an undeclared top-level field a `reject_unknown_fields=true` run would refuse is priced here. An Anthropic Messages model is the exception: its schema allows undeclared fields, so a `reject_unknown_fields=true` run refuses none, and the estimate prices only the undeclared top-level fields `tools`, `tool_choice`, `temperature`, `top_p`, `top_k`, `stop_sequences`, `metadata`, `thinking` and an empty `mcp_servers`. Any other undeclared top-level field makes it answer `unknown` with `reason: not_quotable`.
-- `pricing_as_of` is when the rate card was fetched, or for an `estimated` quote taken from the vendor's published price list, the date that list was read (when neither was read, it is when the estimate was answered). It is not a price lock: the run is charged at the rates in force when it is rated.
+- `pricing_as_of` is when the rate card was fetched. An `estimated` quote taken from the vendor's published price list carries the date that list was read. An `exact` or `estimated` quote whose billed quantity is predicted from a published vendor schedule carries that schedule's read date when it is older than the card. When neither a card nor a schedule was read, it is when the estimate was answered. It is not a price lock: the run is charged at the rates in force when it is rated.
 - `/estimate` quotes the leg `model_provider` names, or the native provider. It takes no `fallback_provider` or `strict_mode` parameter. A synchronous run whose fallback retry lands on another provider is charged at that provider's rate, which the estimate did not quote; `X-Comfy-Router-Fallback-Provider` on the run's response says it happened. A queued submit has no fallback, so its `estimate` quotes the leg that runs.
 - `/estimate` and a queued submit's `estimate` can price slightly different documents for the same body on a `model_provider` leg, or for a body that runs on a sibling operation. The submit's `estimate` is priced on the document the run sends.
 - `/estimate` always quotes the platform charge. A run on your own provider key is not billed by Comfy, but `/estimate` does not detect that and still returns a figure; only a queued submit answers `reason: byok`.
@@ -992,7 +992,7 @@ A model can return asset URLs, inline bytes, or both. The providers below copy s
 | `minimax/*` | the finished video | 12 hours |
 | `xai/*` | every generated image, and the finished video | 24 hours |
 
-These lifetimes start when the URL is signed, not when you open it. Cached or replayed URLs can have less time remaining; replay does not renew them. Download the asset promptly. Only the assets named in each row are copied: `byteplus/seedream-*` and `byteplus/seededit-*` images are not covered by the BytePlus video row.
+These lifetimes start when the URL is signed, not when you open it. Cached or replayed URLs can have less time remaining; replay does not renew them. An `Idempotency-Key` record can outlive the URLs it holds, so from an hour before such a URL expires a same-key retry is answered `409 invalid_input` instead of the recorded response; a response that was already that close to expiry when it was served stays replayable for 5 minutes. Download the asset promptly. Only the assets named in each row are copied: `byteplus/seedream-*` and `byteplus/seededit-*` images are not covered by the BytePlus video row.
 
 Copying is performed by the model's native provider only. A run served by an alternate provider — `model_provider` naming one, or a native run whose fallback retry landed on one (disclosed by `X-Comfy-Router-Fallback-Provider`) — returns that provider's own URLs, which follow the provider's expiry and are not copied. `model_provider` values `default`, `comfy` and `comfyui` name no alternate provider, so those runs are native and copied as above, as is a `model_provider` run whose fallback retry landed back on the native provider.
 
@@ -1099,7 +1099,7 @@ A pre-flight estimate of what this call will cost, in US dollars, priced before
 </ResponseField>
 
 <ResponseField name="pricing_as_of" type="string" required>
-  When the rate card this quote was priced from was fetched. The estimate is not a price lock: the run is charged at the rates in force when it is rated, not at this instant. An `estimated` quote taken from the vendor's published price list (a model billed the cost the vendor reports) carries the date that list was read instead; one whose billed quantity is predicted from that list and then priced from the rate card (a model billed the credits its vendor reports) carries the older of the two dates. Whenever no rate card was read (`reason: pricing_unavailable`, `byok`, or a `not_quotable` call that cannot be priced from any card), this is when the quote was answered.
+  When the rate card this quote was priced from was fetched. The estimate is not a price lock: the run is charged at the rates in force when it is rated, not at this instant. An `estimated` quote taken from the vendor's published price list (a model billed the cost the vendor reports) carries the date that list was read instead. A quote whose billed quantity is predicted from a published vendor schedule and then priced from the rate card carries the older of that schedule's read date and the card's fetch time, for an `exact` answer and an `estimated` one alike (a model billed the credits its vendor reports is the `estimated` case). Whenever no rate card was read (`reason: pricing_unavailable`, `byok`, or a `not_quotable` call that cannot be priced from any card), this is when the quote was answered.
 
   date-time
 </ResponseField>
diff --git a/es/tutorials/video/minimax/minimax-h3-native.mdx b/es/tutorials/video/minimax/minimax-h3-native.mdx
index 1c4348a169..bc76d51183 100644
--- a/es/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/es/tutorials/video/minimax/minimax-h3-native.mdx
@@ -2,14 +2,14 @@
 title: "Flujos de trabajo de MiniMax H3 en ComfyUI: texto a video, imagen a video y referencia a video"
 description: "Ejemplos paso a paso de flujos nativos de MiniMax H3 en ComfyUI, con anclaje de guías y máscaras de ruido latente para un control más fino del video."
 sidebarTitle: "Flujos nativos"
-translationSourceHash: 2a225f15
+translationSourceHash: b16dd815
 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx
 translationBlockHashes:
   "_intro": dde651e5
   "MiniMax H3 Text to Video (T2V)": 42ec8cfb
   "MiniMax H3 Image to Video (I2V)": 2a06e3fa
   "MiniMax H3 Reference to Video (R2V)": 961829df
-  "Advanced workflows with the native nodes": bbaed645
+  "Advanced workflows with the native nodes": d423a0db
 ---
 Esta página cubre las tres plantillas base de MiniMax H3 y las técnicas avanzadas con nodos nativos. Para conocer las capacidades y el licenciamiento, consulta la [descripción general de MiniMax H3](/es/tutorials/video/minimax/minimax-h3).
 
@@ -260,29 +260,31 @@ Conecta una máscara a `denoise_mask` del sampler: **0** = mantener, **1** = reg
 
 ### Clips largos con ventanas de contexto (#17425)
 
-[`MiniMaxH3ContextWindows`](https://github.com/Comfy-Org/ComfyUI/pull/17425) muestrea un clip largo en ventanas superpuestas, así que cada paso mantiene una ventana en memoria en lugar del latente completo. Conéctalo entre `UNETLoader` y `BasicGuider` (model a la entrada, model a la salida), el mismo sitio que cualquier otro parche de modelo.
+[`MiniMaxH3ContextWindows`](https://github.com/Comfy-Org/ComfyUI/pull/17425) muestrea un clip largo en ventanas superpuestas, así que cada paso ejecuta el modelo sobre una ventana en lugar del clip completo. Conéctalo entre el interruptor del LoRA turbo (`ComfySwitchNode`) y `BasicGuider`, con model a la entrada y model a la salida. `BasicScheduler` sigue usando la salida del interruptor y no necesita el parche.
 
 | Control | Significado | Valor por defecto |
 | --- | --- | --- |
-| `context_length` | Tamaño de la ventana en frames, **17n+5** | **124**, la longitud del clip por defecto |
+| `context_length` | Tamaño de la ventana en frames, **17n+5** | **124**, los 5 s de la plantilla ajustados a la rejilla |
 | `context_overlap` | Frames que comparten ventanas contiguas, **17n+5** | **39** |
-| `fuse_method` | Cómo se combinan las ventanas superpuestas | `pyramid` |
+| `fuse_method` | Cómo se combinan las ventanas superpuestas: `pyramid`, `relative`, `flat`, `overlap-linear` | `pyramid` |
 | `freenoise` | Mezcla el ruido entre ventanas para unirlas | Activado |
 
-Las ventanas empiezan en los bloques de 17 frames del modelo, de modo que video y audio siguen alineados al cruzar un límite. El nodo mantiene `context_overlap` por debajo de la longitud de la ventana. Con el valor por defecto **124** el clip de 5 segundos cabe en una sola ventana, así que el nodo solo cambia el muestreo cuando el clip supera la ventana: baja `context_length` para que quepa un clip más largo en la misma tarjeta, súbelo para dar más contexto a cada paso a cambio de más memoria.
+Las ventanas empiezan en los bloques de 17 frames del modelo, de modo que video y audio siguen alineados al cruzar un límite. `context_overlap` tiene como tope la longitud de la ventana menos **17** frames, y un valor mayor se recorta sin aviso. Con el valor por defecto **124** un clip de 5 s ocupa una sola ventana, y la primera longitud que abre una segunda ventana es **141** frames. Baja `context_length` para reducir la memoria de cada paso. Súbelo para dar más contexto a cada paso, con un coste mayor.
+
+La memoria baja por paso, no en total. El latente completo, los búferes que fusionan las ventanas y el decodificado siguen creciendo con la longitud del clip, así que un clip que se queda sin memoria al decodificar sigue igual.
 
 El muestreo por ventanas trae dos límites:
 
-- **El prompt se ejecuta completo dentro de cada ventana.** Los tiempos de plano se leen por ventana, no sobre el clip entero, así que un corte con marca de tiempo vuelve a dispararse en cada una. Describe una escena continua y deja fuera los cortes con tiempo.
+- **Cada ventana recibe el prompt completo.** Describe una escena continua y deja fuera los cortes con tiempo: un desglose de planos con marcas de tiempo da por hecho que el clip se muestrea de una vez.
 - **Las ventanas se fusionan, no se empalman.** El contenido que debe mantenerse idéntico al cruzar un límite (texto en pantalla, detalle del decorado, una cara) puede derivar. Mantén estables la cámara y el sujeto.
 
-Por eso las ventanas de contexto encajan en una escena sin cortes: un push lento, un paneo o un segundo pase de escalado sobre un clip terminado. Para un clip que cambia de plano, genera los planos por separado y únelos.
+Por eso las ventanas de contexto encajan en una escena sin cortes: un push lento, un paneo o un plano sin corte que supera la ventana. Para un clip que cambia de plano, genera los planos por separado y únelos.
 
 ### Generar primero la banda sonora
 
 El audio y el video se desruidizan juntos, así que una programación corta o un LoRA turbo suavizan la pista junto con la imagen. Dos pasadas los separan: fija el audio en un lienzo pequeño y luego condiciona el video con él.
 
-1. **Pasada de audio barata**: Ejecuta una pasada de texto a video a **32×32** con `Boolean (Enable Lightning LoRA)` desactivado, **40–50** pasos y la duración que necesita el plano final. Los frames latentes de audio siguen la longitud del clip, no el tamaño del lienzo, así que esta pasada genera una pista de longitud completa por mucho menos que un render a resolución completa. Desconecta el Resolution Selector y escribe el tamaño en sus destinos `width` y `height`; el selector llega como mínimo a **0.1** megapíxeles. Los pasos están en el `PrimitiveInt` ajustado a **20** dentro del subgrafo (el que está en **6** es el recuento de pasos turbo). De esta pasada solo se conserva el audio.
+1. **Pasada de audio barata**: Ejecuta una pasada de texto a video a **32×32** con `Boolean (Enable Lightning LoRA)` desactivado, **40–50** pasos y la duración que necesita el plano final. Los frames latentes de audio siguen la longitud del clip, no el tamaño del lienzo, así que esta pasada genera una pista de longitud completa por mucho menos que un render a resolución completa. Desconecta el Resolution Selector y escribe el tamaño en sus destinos `width` y `height`; el selector llega como mínimo a **0.1** megapíxeles. Los pasos están en el `PrimitiveInt` ajustado a **20** dentro del subgrafo. El otro `PrimitiveInt` es el recuento de pasos turbo, que el control del subgrafo deja en **8**. De esta pasada solo se conserva el audio.
 2. **Guardar la pista**: El subgrafo decodifica el audio dentro y expone solo el video, así que añade un nodo `SaveAudioAdvanced` dentro del subgrafo y conecta `VAEDecodeAudio` a él.
 3. **Condicionar la segunda pasada**: En el flujo R2V añade `LoadAudio`, conéctalo a `ref_audio` y mantén `audio_vae` conectado. Cita la pista como `<Audio 1>` e indica la relación: `reference` conserva timbre, ritmo y dicción sin copiar la señal, `fully_copy` la convierte en la pista final completa, `partially_copy` conserva parte de la línea de tiempo o capas seleccionadas, y `weak_reference` solo mantiene una similitud amplia. La segunda pasada parte así de una voz y un ritmo ya resueltos en lugar de inventar el habla desde cero. Para usar el mismo archivo, conecta `LoadAudio` a `CreateVideo` en vez de confiar en el prompt.
 
diff --git a/ja/tutorials/video/minimax/minimax-h3-native.mdx b/ja/tutorials/video/minimax/minimax-h3-native.mdx
index b6bbc7fd2a..c07961e714 100644
--- a/ja/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/ja/tutorials/video/minimax/minimax-h3-native.mdx
@@ -2,14 +2,14 @@
 title: "ComfyUI MiniMax H3 テキストから動画・画像から動画・参照から動画のワークフロー"
 description: "ComfyUI における MiniMax H3 ネイティブワークフローの手順例。ガイドのアンカー固定や latent ノイズマスクで生成動画をより細かく制御できます。"
 sidebarTitle: "ネイティブワークフロー"
-translationSourceHash: 2a225f15
+translationSourceHash: b16dd815
 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx
 translationBlockHashes:
   "_intro": dde651e5
   "MiniMax H3 Text to Video (T2V)": 42ec8cfb
   "MiniMax H3 Image to Video (I2V)": 2a06e3fa
   "MiniMax H3 Reference to Video (R2V)": 961829df
-  "Advanced workflows with the native nodes": bbaed645
+  "Advanced workflows with the native nodes": d423a0db
 ---
 
 このページは 3 つの基本 MiniMax H3 テンプレートとネイティブノードの高度なテクニックを扱います。機能とライセンスは [MiniMax H3 概要](/ja/tutorials/video/minimax/minimax-h3) を参照してください。
@@ -261,29 +261,31 @@ H3 ノードの `positive` と latent を `MiniMaxH3AddGuide` に接続:
 
 ### コンテキストウィンドウで長いクリップを生成 (#17425)
 
-[`MiniMaxH3ContextWindows`](https://github.com/Comfy-Org/ComfyUI/pull/17425) は長いクリップを重なり合うウィンドウに分割してサンプリングし、各ステップで全体ではなく 1 ウィンドウ分だけをメモリに載せます。`UNETLoader` と `BasicGuider` の間（model 入力、model 出力）に接続します。他の model パッチと同じ位置です。
+[`MiniMaxH3ContextWindows`](https://github.com/Comfy-Org/ComfyUI/pull/17425) は長いクリップを重なり合うウィンドウに分けてサンプリングし、各ステップではクリップ全体ではなく 1 ウィンドウだけでモデルを動かします。turbo LoRA のスイッチ（`ComfySwitchNode`）と `BasicGuider` の間に、model 入力・model 出力で接続します。`BasicScheduler` はスイッチの出力のままでよく、このパッチは不要です。
 
 | 項目 | 意味 | デフォルト |
 | --- | --- | --- |
-| `context_length` | ウィンドウ長（フレーム）、**17n+5** | **124**（デフォルトのクリップ長） |
+| `context_length` | ウィンドウ長（フレーム）、**17n+5** | **124**（テンプレートの 5 秒をグリッドに合わせた値） |
 | `context_overlap` | 隣接ウィンドウが共有するフレーム数、**17n+5** | **39** |
-| `fuse_method` | 重なったウィンドウの融合方法 | `pyramid` |
+| `fuse_method` | 重なったウィンドウの融合方法: `pyramid`、`relative`、`flat`、`overlap-linear` | `pyramid` |
 | `freenoise` | ウィンドウ間でノイズをシャッフルしてつなぐ | 有効 |
 
-ウィンドウはモデルの 17 フレーム塊の境界から始まるため、動画と音声は境界をまたいでも揃います。ノードは `context_overlap` をウィンドウ長未満に抑えます。デフォルトの **124** では 5 秒のクリップ全体が 1 ウィンドウに入るので、クリップがウィンドウを超えてから初めて効果が出ます。`context_length` を下げると同じ GPU で長いクリップが入り、上げると各ステップの前後文脈が増える代わりにメモリを多く使います。
+ウィンドウはモデルの 17 フレーム塊の境界から始まるため、動画と音声は境界をまたいでも揃います。`context_overlap` の上限はウィンドウ長から **17** フレームを引いた値で、それより大きく設定しても警告なしに切り詰められます。デフォルトの **124** では 5 秒のクリップが 1 ウィンドウに収まり、2 つ目のウィンドウが開く最小の長さは **141** フレームです。`context_length` を下げると 1 ステップあたりのメモリが減ります。上げると 1 ステップあたりの前後文脈が増え、コストも上がります。
+
+減るのは 1 ステップあたりのメモリで、総量ではありません。全体の latent、ウィンドウ融合用のバッファ、デコードはクリップ長に比例したままなので、デコードで OOM になるクリップはそのまま OOM になります。
 
 ウィンドウ分割には 2 つの制約があります:
 
-- **各ウィンドウでプロンプト全体が実行されます。** ショットの時刻はクリップ全体ではなくウィンドウ基準で読まれるため、時刻指定のカットは各ウィンドウで再度発生します。1 つの連続したシーンとして書き、時刻指定のカットは入れません。
+- **どのウィンドウにもプロンプト全体が渡ります。** 1 つの連続したシーンとして書き、時刻指定のカットは入れません。時刻付きのカット割りは、クリップ全体を 1 回でサンプリングする前提の書き方です。
 - **ウィンドウは融合されるのであって連結ではありません。** 境界をまたいで厳密に保つ必要がある内容（画面内の文字、セットの細部、顔）はドリフトし得ます。カメラと被写体は安定させます。
 
-そのためコンテキストウィンドウは、カットのない連続したシーン（ゆっくりしたプッシュイン、パン、完成クリップの 2 パスアップスケール）に向きます。ショットを切り替えるクリップは、ショットごとに生成してからつなぎます。
+そのためコンテキストウィンドウは、カットのない連続したシーン（ゆっくりしたプッシュイン、パン、ウィンドウ長を超えるワンカット）に向きます。ショットを切り替えるクリップは、ショットごとに生成してからつなぎます。
 
 ### 音声を先に作る
 
 音声と動画は同じデノイズで生成されるため、短いスケジュールや turbo LoRA は映像と一緒に音声も甘くします。2 パスに分ければ切り離せます。まず小さいキャンバスで音声を固め、その音声を条件に映像を作ります。
 
-1. **安い音声パス**: **32×32** で `Boolean (Enable Lightning LoRA)` をオフ、**40–50** ステップ、最終カットに必要な長さでテキストから動画のパスを 1 回実行します。音声 latent のフレーム数はクリップ長に従い、キャンバスサイズには依存しないため、このパスはフル解像度のレンダーよりはるかに安いコストでフル尺の音声を作れます。Resolution Selector を切り離し、対応する `width` と `height` にサイズを直接入力します。セレクターの下限は **0.1** メガピクセルです。ステップ数はサブグラフ内で **20** に設定された `PrimitiveInt` にあります（**6** のほうは turbo のステップ数です）。このパスでは音声だけを使います。
+1. **安い音声パス**: **32×32** で `Boolean (Enable Lightning LoRA)` をオフ、**40–50** ステップ、最終カットに必要な長さでテキストから動画のパスを 1 回実行します。音声 latent のフレーム数はクリップ長に従い、キャンバスサイズには依存しないため、このパスはフル解像度のレンダーよりはるかに安いコストでフル尺の音声を作れます。Resolution Selector を切り離し、対応する `width` と `height` にサイズを直接入力します。セレクターの下限は **0.1** メガピクセルです。ステップ数はサブグラフ内で **20** に設定された `PrimitiveInt` にあります。もう 1 つの `PrimitiveInt` が turbo のステップ数で、サブグラフのウィジェットでは **8** になっています。このパスでは音声だけを使います。
 2. **トラックを保存**: サブグラフは内部で音声をデコードし、映像だけを出力します。サブグラフ内に `SaveAudioAdvanced` を追加し、`VAEDecodeAudio` を接続してください。
 3. **2 パス目を条件づける**: R2V ワークフローに `LoadAudio` を追加して `ref_audio` に接続し、`audio_vae` は接続したままにします。プロンプトでは `<Audio 1>` として参照し、関係を明記します。`reference` は信号をコピーせず音色・リズム・話し方だけを参照し、`fully_copy` は完全な最終トラックとして扱い、`partially_copy` はタイムラインの一部または一部の音声レイヤーだけをコピーし、`weak_reference` はカテゴリや雰囲気の広い類似だけを保ちます。2 パス目はゼロから音声を作るのではなく、確定した声とタイミングから始まります。まったく同じファイルを使う場合は、プロンプトに頼らず `LoadAudio` を `CreateVideo` に接続します。
 
diff --git a/ko/tutorials/video/minimax/minimax-h3-native.mdx b/ko/tutorials/video/minimax/minimax-h3-native.mdx
index a72c906344..6f373141df 100644
--- a/ko/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/ko/tutorials/video/minimax/minimax-h3-native.mdx
@@ -2,14 +2,14 @@
 title: "ComfyUI MiniMax H3 텍스트·이미지·레퍼런스 기반 비디오 워크플로"
 description: "ComfyUI의 MiniMax H3 네이티브 워크플로 단계별 예제입니다. 가이드 고정과 잠재 노이즈 마스크로 생성 비디오를 더 세밀하게 제어할 수 있습니다."
 sidebarTitle: "네이티브 워크플로"
-translationSourceHash: 2a225f15
+translationSourceHash: b16dd815
 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx
 translationBlockHashes:
   "_intro": dde651e5
   "MiniMax H3 Text to Video (T2V)": 42ec8cfb
   "MiniMax H3 Image to Video (I2V)": 2a06e3fa
   "MiniMax H3 Reference to Video (R2V)": 961829df
-  "Advanced workflows with the native nodes": bbaed645
+  "Advanced workflows with the native nodes": d423a0db
 ---
 
 이 페이지는 세 가지 기본 MiniMax H3 템플릿과 네이티브 노드 고급 기법을 다룹니다. 기능과 라이선스는 [MiniMax H3 개요](/ko/tutorials/video/minimax/minimax-h3)를 참조하세요.
@@ -261,29 +261,31 @@ H3 노드 `positive`·latent를 `MiniMaxH3AddGuide`에 연결:
 
 ### 컨텍스트 윈도우로 긴 클립 생성 (#17425)
 
-[`MiniMaxH3ContextWindows`](https://github.com/Comfy-Org/ComfyUI/pull/17425)는 긴 클립을 겹치는 윈도우로 나눠 샘플링합니다. 각 단계에서 전체 latent 대신 윈도우 하나만 메모리에 올립니다. `UNETLoader`와 `BasicGuider` 사이(model 입력, model 출력)에 연결합니다. 다른 model 패치와 같은 자리입니다.
+[`MiniMaxH3ContextWindows`](https://github.com/Comfy-Org/ComfyUI/pull/17425)는 긴 클립을 겹치는 윈도우로 나눠 샘플링합니다. 각 단계에서 클립 전체가 아니라 윈도우 하나로 모델을 돌립니다. turbo LoRA 스위치(`ComfySwitchNode`)와 `BasicGuider` 사이에 model 입력, model 출력으로 연결합니다. `BasicScheduler`는 스위치 출력을 그대로 쓰면 되고 이 패치는 필요 없습니다.
 
 | 항목 | 의미 | 기본값 |
 | --- | --- | --- |
-| `context_length` | 윈도우 길이(프레임), **17n+5** | **124**, 기본 클립 길이 |
+| `context_length` | 윈도우 길이(프레임), **17n+5** | **124**, 템플릿의 5초를 그리드에 맞춘 값 |
 | `context_overlap` | 인접 윈도우가 공유하는 프레임 수, **17n+5** | **39** |
-| `fuse_method` | 겹친 윈도우를 합치는 방식 | `pyramid` |
+| `fuse_method` | 겹친 윈도우를 합치는 방식: `pyramid`, `relative`, `flat`, `overlap-linear` | `pyramid` |
 | `freenoise` | 윈도우 사이 노이즈를 섞어 이어 붙임 | 켜짐 |
 
-윈도우는 모델의 17프레임 덩어리 경계에서 시작하므로 영상과 오디오가 경계를 넘어도 어긋나지 않습니다. 노드는 `context_overlap`을 윈도우 길이 아래로 제한합니다. 기본값 **124**에서는 5초 클립 전체가 윈도우 하나에 들어가므로, 클립이 윈도우보다 길어져야 효과가 생깁니다. `context_length`를 낮추면 같은 GPU에서 더 긴 클립이 들어가고, 높이면 각 단계가 더 많은 앞뒤 맥락을 보는 대신 메모리를 더 씁니다.
+윈도우는 모델의 17프레임 덩어리 경계에서 시작하므로 영상과 오디오가 경계를 넘어도 어긋나지 않습니다. `context_overlap` 상한은 윈도우 길이에서 **17**프레임을 뺀 값이고, 더 크게 넣으면 경고 없이 잘립니다. 기본값 **124**에서는 5초 클립이 윈도우 하나에 들어가고, 두 번째 윈도우가 열리는 최소 길이는 **141**프레임입니다. `context_length`를 낮추면 단계마다 쓰는 메모리가 줄어듭니다. 높이면 단계마다 더 많은 앞뒤 맥락을 보는 대신 비용이 커집니다.
+
+줄어드는 것은 단계당 메모리이지 총량이 아닙니다. 전체 latent, 윈도우를 합치는 버퍼, 디코드는 클립 길이에 비례하므로, 디코드에서 메모리가 부족한 클립은 그대로 부족합니다.
 
 윈도우 분할에는 제약이 두 가지 있습니다:
 
-- **모든 윈도우에서 프롬프트 전체가 실행됩니다.** 샷 시각은 클립 전체가 아니라 윈도우 기준으로 읽히므로, 시각을 지정한 컷은 윈도우마다 다시 발생합니다. 하나의 연속된 장면으로 쓰고 시각 지정 컷은 넣지 않습니다.
+- **모든 윈도우가 프롬프트 전체를 받습니다.** 하나의 연속된 장면으로 쓰고 시각 지정 컷은 넣지 않습니다. 시각이 붙은 컷 구성은 클립 전체를 한 번에 샘플링한다는 전제로 쓰인 것입니다.
 - **윈도우는 이어 붙이는 게 아니라 합쳐집니다.** 경계를 넘어 정확히 유지해야 하는 내용(화면 속 글자, 세트 디테일, 얼굴)은 밀릴 수 있습니다. 카메라와 피사체는 안정적으로 유지합니다.
 
-그래서 컨텍스트 윈도우는 컷이 없는 연속 장면에 맞습니다. 느린 푸시인, 팬, 완성된 클립의 2패스 업스케일 같은 경우입니다. 샷을 전환하는 클립은 샷별로 생성한 뒤 이어 붙입니다.
+그래서 컨텍스트 윈도우는 컷이 없는 연속 장면에 맞습니다. 느린 푸시인, 팬, 윈도우보다 긴 원테이크 같은 경우입니다. 샷을 전환하는 클립은 샷별로 생성한 뒤 이어 붙입니다.
 
 ### 사운드트랙을 먼저 만들기
 
 오디오와 비디오는 같은 디노이즈에서 함께 생성되므로, 짧은 스케줄이나 turbo LoRA는 영상과 함께 오디오도 무디게 만듭니다. 두 번의 패스로 분리할 수 있습니다. 작은 캔버스에서 오디오를 먼저 확정하고, 그 오디오를 조건으로 영상을 만듭니다.
 
-1. **저렴한 오디오 패스**: **32×32**, `Boolean (Enable Lightning LoRA)` 끔, **40–50** 스텝, 최종 샷에 필요한 길이로 텍스트-투-비디오 패스를 한 번 실행합니다. 오디오 latent 프레임 수는 클립 길이를 따르고 캔버스 크기와 무관하므로, 이 패스는 전체 해상도 렌더보다 훨씬 낮은 비용으로 전체 길이 트랙을 만듭니다. Resolution Selector 연결을 끊고 해당 `width`, `height`에 값을 직접 입력하세요. 선택기의 하한은 **0.1** 메가픽셀입니다. 스텝 수는 서브그래프 안에서 **20**으로 설정된 `PrimitiveInt`에 있습니다(**6**인 쪽은 turbo 스텝 수입니다). 이 패스에서는 오디오만 사용합니다.
+1. **저렴한 오디오 패스**: **32×32**, `Boolean (Enable Lightning LoRA)` 끔, **40–50** 스텝, 최종 샷에 필요한 길이로 텍스트-투-비디오 패스를 한 번 실행합니다. 오디오 latent 프레임 수는 클립 길이를 따르고 캔버스 크기와 무관하므로, 이 패스는 전체 해상도 렌더보다 훨씬 낮은 비용으로 전체 길이 트랙을 만듭니다. Resolution Selector 연결을 끊고 해당 `width`, `height`에 값을 직접 입력하세요. 선택기의 하한은 **0.1** 메가픽셀입니다. 스텝 수는 서브그래프 안에서 **20**으로 설정된 `PrimitiveInt`에 있습니다. 다른 `PrimitiveInt`가 turbo 스텝 수이고, 서브그래프 위젯에서는 **8**로 설정됩니다. 이 패스에서는 오디오만 사용합니다.
 2. **트랙 저장**: 서브그래프는 내부에서 오디오를 디코드하고 비디오만 노출하므로, 서브그래프 안에 `SaveAudioAdvanced`를 추가하고 `VAEDecodeAudio`를 연결하세요.
 3. **두 번째 패스 조건 걸기**: R2V 워크플로에 `LoadAudio`를 추가해 `ref_audio`에 연결하고 `audio_vae`는 연결된 상태로 둡니다. 프롬프트에서 `<Audio 1>`로 참조하고 관계를 명시하세요. `reference`는 신호를 복사하지 않고 음색, 리듬, 말투만 참조하고, `fully_copy`는 완전한 최종 트랙으로 취급하며, `partially_copy`는 타임라인의 일부나 일부 오디오 레이어만 복사하고, `weak_reference`는 범주와 분위기의 넓은 유사성만 유지합니다. 두 번째 패스는 빈 상태에서 음성을 지어내는 대신 확정된 목소리와 타이밍에서 시작합니다. 완전히 같은 파일을 쓰려면 프롬프트에 의존하지 말고 `LoadAudio`를 `CreateVideo`에 연결하세요.
 
diff --git a/router-openapi.yaml b/router-openapi.yaml
index 021df9e08b..e18751da9c 100644
--- a/router-openapi.yaml
+++ b/router-openapi.yaml
@@ -568,7 +568,7 @@ components:
         pricing_as_of:
           type: string
           format: date-time
-          description: 'When the rate card this quote was priced from was fetched. The estimate is not a price lock: the run is charged at the rates in force when it is rated, not at this instant. An `estimated` quote taken from the vendor''s published price list (a model billed the cost the vendor reports) carries the date that list was read instead; one whose billed quantity is predicted from that list and then priced from the rate card (a model billed the credits its vendor reports) carries the older of the two dates. Whenever no rate card was read (`reason: pricing_unavailable`, `byok`, or a `not_quotable` call that cannot be priced from any card), this is when the quote was answered.'
+          description: 'When the rate card this quote was priced from was fetched. The estimate is not a price lock: the run is charged at the rates in force when it is rated, not at this instant. An `estimated` quote taken from the vendor''s published price list (a model billed the cost the vendor reports) carries the date that list was read instead. A quote whose billed quantity is predicted from a published vendor schedule and then priced from the rate card carries the older of that schedule''s read date and the card''s fetch time, for an `exact` answer and an `estimated` one alike (a model billed the credits its vendor reports is the `estimated` case). Whenever no rate card was read (`reason: pricing_unavailable`, `byok`, or a `not_quotable` call that cannot be priced from any card), this is when the quote was answered.'
         reason:
           type: string
           description: 'Present only when `source` is `unknown`: `not_quotable` (Router cannot price this call before it runs - its price depends on what the provider reports back, or the body does not bound a billed quantity), `unpriced_model` (the rate card does not price this request), `byok` (queued submit only: the run uses your own provider key, so Comfy bills nothing for it), or `pricing_unavailable` (queued submit only: the rate card could not be read when the request was admitted; this is transient, and the run is still charged as usual). Read any other value as `not_quotable`.'
diff --git a/router-schemas/runway/gen4_image.json b/router-schemas/runway/gen4_image.json
index b56d4ef59f..f47a8560a8 100644
--- a/router-schemas/runway/gen4_image.json
+++ b/router-schemas/runway/gen4_image.json
@@ -1 +1 @@
-{"openapi":"3.0.2","info":{"title":"runway/gen4_image","description":"The request body Comfy Router accepts for the model \"runway/gen4_image\", and the response body it returns. The INPUT schema is the same schema the server validates a call against before it reaches the provider, so what is published and what is enforced cannot differ. `x-comfy-router-input-modalities` records the media kinds the model is documented to accept; the request schema is shared across sibling models and may admit media kinds this model ignores. The OUTPUT schema describes the provider's native result document exactly as Router returns it: Router does not validate, narrow or re-envelope the response, so the output schema is descriptive rather than enforced, and Comfy owns no output shape of its own.","version":"7e48312b99fa"},"paths":{"/v2/models/runway/gen4_image":{"post":{"operationId":"runRouterModel","summary":"Run runway/gen4_image synchronously.","security":[{"BearerAuth":[]},{"ApiKeyAuth":[]}],"requestBody":{"required":true,"content":{"application/json":{"schema":{"example":{"promptText":"a red circle","ratio":"1024:1024"},"properties":{"model":{"description":"Model to use for generation. On the Comfy Router route `POST /v2/models/runway/{model}` this field is supplied from the path and MUST NOT be sent; on the v1 `POST /proxy/runway/text_to_image` route it is required and constrained to `gen4_image`.","nullable":true,"type":"string"},"promptText":{"description":"Text prompt for the image generation","maxLength":1000,"type":"string"},"ratio":{"$ref":"#/components/schemas/RunwayTextToImageAspectRatioEnum"},"referenceImages":{"description":"Array of reference images to guide the generation","items":{"properties":{"uri":{"description":"A HTTPS URL or data URI containing an encoded image","type":"string"}},"required":["uri"],"type":"object"},"type":"array"}},"required":["promptText","ratio"],"type":"object"}}}},"responses":{"200":{"description":"OK - the model's native JSON output, returned unchanged; the schema describes the provider's terminal result document as Router returns it.","content":{"application/json":{"schema":{"allOf":[{"$ref":"#/components/schemas/RunwayTaskStatusResponse"},{"properties":{"output":{"description":"The finished generation's asset URLs — for `runway/gen4_image` each element is an IMAGE URL. Present once the task terminates successfully; this list IS the result.","items":{"type":"string"},"type":"array"}},"type":"object"}],"description":"Comfy Router output schema for the Runway image model: the terminal `GET /v1/tasks/{id}` document, forwarded unchanged. `runway/gen4_image` is SUBMIT-AND-POLL DESPITE being an image generation — `routerresult/classification.go` records `{provider: runway, endpoint: /text_to_image}` as `ReturnModeSubmitPoll`, sharing the one `GET /proxy/runway/tasks/{task_id}` poll route its video siblings use — because Runway's images go through the same task queue its videos do and are collected from the same status route. Router polls on the caller's behalf, so the body a caller receives is the finished task rather than the `{\"id\": ...}` handle the underlying submit answers with.\nTHE ASSET IS THE `output` LIST ITSELF. `output` is an array of URL STRINGS — not objects with a URL inside — so the leaf a caller reads is `output[]`, and there is no deeper field to key completion off. For this model its element is a GENERATED IMAGE URL. The video models of the same family answer the same document with a video URL in that list; see `RunwayVideoRouterOutput`.\n`output` is absent until the task reaches a terminal state, and a `SUCCEEDED` task carrying an empty `output` is a success that produced nothing — `routerpollstate.classifyRunway` (`FamilyRunway`) draws exactly that line, keying success on a non-empty `output` and answering `success_without_output` otherwise. Through Router that distinction is already settled, because such a task is answered as a Comfy Router error rather than with this document, so a `200` here always carries at least one URL.\nThe URLs are served from Runway's own asset host and are NOT permanent — download them promptly rather than storing them. `status` is Runway's own UPPERCASE vocabulary (`PENDING`, `RUNNING`, `THROTTLED`, `SUCCEEDED`, `FAILED`, `CANCELLED`); `SUCCEEDED` is the one terminal success, `FAILED` and `CANCELLED` are terminal failures, and the classifier upper-cases before comparing. `progress` is populated only while `status` is `RUNNING`, so it is not present on the document Router returns.","example":{"createdAt":"2027-01-01T00:00:00Z","id":"5f7c1b28-9a03-4e61-8d2f-1b2c3d4e5f60","output":["https://example.invalid/runway/gen4_image/generated.png"],"status":"SUCCEEDED"}}}}}}}}},"components":{"schemas":{"RunwayTaskStatusEnum":{"description":"Possible statuses for a Runway task.","enum":["SUCCEEDED","RUNNING","FAILED","PENDING","CANCELLED","THROTTLED"],"type":"string"},"RunwayTaskStatusResponse":{"properties":{"createdAt":{"description":"Task creation timestamp","format":"date-time","type":"string"},"id":{"description":"Task ID","type":"string"},"output":{"description":"Array of the finished task's output asset URLs. This route is shared by every Runway model, so the medium follows the model that was submitted - a video URL for the image_to_video and video_to_video models, an image URL for the text_to_image model.","items":{"type":"string"},"type":"array"},"progress":{"description":"Float value between 0 and 1 representing the progress of the task. Only available if status is RUNNING.","format":"float","maximum":1,"minimum":0,"type":"number"},"status":{"$ref":"#/components/schemas/RunwayTaskStatusEnum"}},"required":["id","status","createdAt"],"type":"object"},"RunwayTextToImageAspectRatioEnum":{"enum":["1920:1080","1080:1920","1024:1024","1360:768","1080:1080","1168:880","1440:1080","1080:1440","1808:768","2112:912"],"type":"string"}},"securitySchemes":{"ApiKeyAuth":{"type":"apiKey","in":"header","name":"X-API-Key"},"BearerAuth":{"type":"http","scheme":"bearer","bearerFormat":"JWT"}}},"x-comfy-router-model-id":"runway/gen4_image","x-comfy-input-schema-authored":true,"x-comfy-output-schema-authored":true,"x-comfy-router-input-modalities":["text","image"],"x-comfy-router-output-modalities":["image"]}
+{"openapi":"3.0.2","info":{"title":"runway/gen4_image","description":"The request body Comfy Router accepts for the model \"runway/gen4_image\", and the response body it returns. The INPUT schema is the same schema the server validates a call against before it reaches the provider, so what is published and what is enforced cannot differ. `x-comfy-router-input-modalities` records the media kinds the model is documented to accept; the request schema is shared across sibling models and may admit media kinds this model ignores. The OUTPUT schema describes the provider's native result document exactly as Router returns it: Router does not validate, narrow or re-envelope the response, so the output schema is descriptive rather than enforced, and Comfy owns no output shape of its own.","version":"4f5b28e16448"},"paths":{"/v2/models/runway/gen4_image":{"post":{"operationId":"runRouterModel","summary":"Run runway/gen4_image synchronously.","security":[{"BearerAuth":[]},{"ApiKeyAuth":[]}],"requestBody":{"required":true,"content":{"application/json":{"schema":{"example":{"promptText":"a red circle","ratio":"1024:1024"},"properties":{"model":{"description":"Model to use for generation. On the Comfy Router route `POST /v2/models/runway/{model}` this field is supplied from the path and MUST NOT be sent; on the v1 `POST /proxy/runway/text_to_image` route it is required and constrained to `gen4_image`.","nullable":true,"type":"string"},"promptText":{"description":"Text prompt for the image generation","maxLength":1000,"type":"string"},"ratio":{"$ref":"#/components/schemas/RunwayTextToImageAspectRatioEnum"},"referenceImages":{"description":"Array of reference images to guide the generation","items":{"properties":{"uri":{"description":"A HTTPS URL, Runway upload URI (runway://) or data URI containing an encoded image","type":"string"}},"required":["uri"],"type":"object"},"maxItems":3,"type":"array"}},"required":["promptText","ratio"],"type":"object"}}}},"responses":{"200":{"description":"OK - the model's native JSON output, returned unchanged; the schema describes the provider's terminal result document as Router returns it.","content":{"application/json":{"schema":{"allOf":[{"$ref":"#/components/schemas/RunwayTaskStatusResponse"},{"properties":{"output":{"description":"The finished generation's asset URLs — for `runway/gen4_image` each element is an IMAGE URL. Present once the task terminates successfully; this list IS the result.","items":{"type":"string"},"type":"array"}},"type":"object"}],"description":"Comfy Router output schema for the Runway image model: the terminal `GET /v1/tasks/{id}` document, forwarded unchanged. `runway/gen4_image` is SUBMIT-AND-POLL DESPITE being an image generation — `routerresult/classification.go` records `{provider: runway, endpoint: /text_to_image}` as `ReturnModeSubmitPoll`, sharing the one `GET /proxy/runway/tasks/{task_id}` poll route its video siblings use — because Runway's images go through the same task queue its videos do and are collected from the same status route. Router polls on the caller's behalf, so the body a caller receives is the finished task rather than the `{\"id\": ...}` handle the underlying submit answers with.\nTHE ASSET IS THE `output` LIST ITSELF. `output` is an array of URL STRINGS — not objects with a URL inside — so the leaf a caller reads is `output[]`, and there is no deeper field to key completion off. For this model its element is a GENERATED IMAGE URL. The video models of the same family answer the same document with a video URL in that list; see `RunwayVideoRouterOutput`.\n`output` is absent until the task reaches a terminal state, and a `SUCCEEDED` task carrying an empty `output` is a success that produced nothing — `routerpollstate.classifyRunway` (`FamilyRunway`) draws exactly that line, keying success on a non-empty `output` and answering `success_without_output` otherwise. Through Router that distinction is already settled, because such a task is answered as a Comfy Router error rather than with this document, so a `200` here always carries at least one URL.\nThe URLs are served from Runway's own asset host and are NOT permanent — download them promptly rather than storing them. `status` is Runway's own UPPERCASE vocabulary (`PENDING`, `RUNNING`, `THROTTLED`, `SUCCEEDED`, `FAILED`, `CANCELLED`); `SUCCEEDED` is the one terminal success, `FAILED` and `CANCELLED` are terminal failures, and the classifier upper-cases before comparing. `progress` is populated only while `status` is `RUNNING`, so it is not present on the document Router returns.","example":{"createdAt":"2027-01-01T00:00:00Z","id":"5f7c1b28-9a03-4e61-8d2f-1b2c3d4e5f60","output":["https://example.invalid/runway/gen4_image/generated.png"],"status":"SUCCEEDED"}}}}}}}}},"components":{"schemas":{"RunwayTaskStatusEnum":{"description":"Possible statuses for a Runway task.","enum":["SUCCEEDED","RUNNING","FAILED","PENDING","CANCELLED","THROTTLED"],"type":"string"},"RunwayTaskStatusResponse":{"properties":{"createdAt":{"description":"Task creation timestamp","format":"date-time","type":"string"},"id":{"description":"Task ID","type":"string"},"output":{"description":"Array of the finished task's output asset URLs. This route is shared by every Runway model, so the medium follows the model that was submitted - a video URL for the image_to_video and video_to_video models, an image URL for the text_to_image model.","items":{"type":"string"},"type":"array"},"progress":{"description":"Float value between 0 and 1 representing the progress of the task. Only available if status is RUNNING.","format":"float","maximum":1,"minimum":0,"type":"number"},"status":{"$ref":"#/components/schemas/RunwayTaskStatusEnum"}},"required":["id","status","createdAt"],"type":"object"},"RunwayTextToImageAspectRatioEnum":{"enum":["1920:1080","1080:1920","1024:1024","1360:768","1080:1080","1168:880","1440:1080","1080:1440","1808:768","2112:912"],"type":"string"}},"securitySchemes":{"ApiKeyAuth":{"type":"apiKey","in":"header","name":"X-API-Key"},"BearerAuth":{"type":"http","scheme":"bearer","bearerFormat":"JWT"}}},"x-comfy-router-model-id":"runway/gen4_image","x-comfy-input-schema-authored":true,"x-comfy-output-schema-authored":true,"x-comfy-router-input-modalities":["text","image"],"x-comfy-router-output-modalities":["image"]}
diff --git a/router-schemas/xai/grok-imagine-image-2.0.json b/router-schemas/xai/grok-imagine-image-2.0.json
index c4d26e6417..abda42a166 100644
--- a/router-schemas/xai/grok-imagine-image-2.0.json
+++ b/router-schemas/xai/grok-imagine-image-2.0.json
@@ -1 +1 @@
-{"openapi":"3.0.2","info":{"title":"xai/grok-imagine-image-2.0","description":"The request body Comfy Router accepts for the model \"xai/grok-imagine-image-2.0\", and the response body it returns. The INPUT schema is the same schema the server validates a call against before it reaches the provider, so what is published and what is enforced cannot differ. `x-comfy-router-input-modalities` records the media kinds the model is documented to accept; the request schema is shared across sibling models and may admit media kinds this model ignores. The OUTPUT schema describes the provider's native result document exactly as Router returns it: Router does not validate, narrow or re-envelope the response, so the output schema is descriptive rather than enforced, and Comfy owns no output shape of its own.","version":"755fafabd68d"},"paths":{"/v2/models/xai/grok-imagine-image-2.0":{"post":{"operationId":"runRouterModel","summary":"Run xai/grok-imagine-image-2.0 synchronously.","security":[{"BearerAuth":[]},{"ApiKeyAuth":[]}],"requestBody":{"required":true,"content":{"application/json":{"schema":{"description":"Request body for xAI Grok Imagine image generation","example":{"n":1,"prompt":"A single red maple leaf on a plain white background."},"properties":{"aspect_ratio":{"default":"auto","description":"Aspect ratio of the generated image. Defaults to auto for automatically selecting the best ratio for the prompt.","enum":["1:1","3:4","4:3","9:16","16:9","2:3","3:2","9:19.5","19.5:9","9:20","20:9","1:2","2:1","auto"],"type":"string"},"model":{"default":"grok-imagine-image","description":"Model to be used. Supported: grok-imagine-image (default), grok-imagine-image-pro, grok-imagine-image-quality, grok-imagine-image-2.0. Deprecated -beta ids are aliased to their GA model.","nullable":true,"type":"string"},"n":{"default":1,"description":"Number of images to be generated","maximum":10,"minimum":1,"type":"integer"},"prompt":{"description":"Prompt for image generation","type":"string"},"quality":{"description":"Quality of the output image. For grok-imagine-image-2.0 this selects the price tier (low/medium; medium is the default); other models currently ignore it.","enum":["low","medium","high"],"type":"string"},"resolution":{"default":"1k","description":"Resolution of the generated image. Defaults to 1k.","enum":["1k","2k"],"type":"string"},"response_format":{"default":"url","description":"Response format to return the image in. Can be url or b64_json. Comfy coerces this to `url` on the outbound request — on the Comfy Router dispatch (`POST /v2/models/xai/{model}`) and on this `/proxy/` route too — because image results are served as re-hosted URLs either way. A value IN the enum is therefore accepted and ignored; send `url` or `b64_json` or omit the field, and the answer is the same. A value OUTSIDE the enum — excluding the UNSET spellings each surface recognises, described next — is REJECTED rather than coerced on BOTH surfaces, and only the shape of the refusal differs: this `/proxy/` route answers a `400` at ingress, and the `POST /v2/models/xai/{model}` Router dispatch answers the per-field `422` it owes the enum published here, before any partner call is made. The two surfaces differ on what counts as UNSET. On this `/proxy/` route an omitted field, an explicit `null` and `\"\"` all mean unset and are coerced like any accepted value. The Router dispatch enforces the enum exactly as published, so `null` and `\"\"` are non-members there like any other: OMIT the field to leave it unset on that surface.","enum":["url","b64_json"],"type":"string"},"size":{"description":"Size of the image (not supported)","type":"string"},"style":{"description":"Style of the image (not supported)","type":"string"},"user":{"description":"A unique identifier representing your end-user, which can help xAI to monitor and detect abuse","type":"string"}},"required":["prompt"],"type":"object"}}}},"responses":{"200":{"description":"OK - the model's native JSON output, returned unchanged; the schema describes the provider's terminal result document as Router returns it.","content":{"application/json":{"schema":{"allOf":[{"$ref":"#/components/schemas/XAIImageGenerationResponse"}],"description":"Comfy Router output schema for the xAI Grok Imagine image models: the `POST /v1/images/generations` document xAI returns, forwarded unchanged EXCEPT for the re-hosted `data[]` entries described below - every other field is xAI's own. This family is DIRECT-RETURN — `routerresult/classification.go` records `{provider: xai, endpoint: /v1/images/generations}` as `ReturnModeDirect` with no poll route — so the body a caller receives is the finished generation on the original call rather than a task handle.\nThe generated images are in `data`, and on THIS surface an entry is addressed by its `url`. Router coerces the outbound `response_format` to `url` for this family and re-hosts every image it can fetch onto Comfy storage, replacing xAI's URL with a signed Comfy one and clearing `b64_json`, so a request that asked for `b64_json` is answered exactly as a `url` request is. `mime_type` names the encoding. The `/proxy/` image routes coerce `response_format` to `url` the same way, so they do not preserve a `b64_json` request either.\nDurability is per entry, not per response: an image whose re-host fails keeps xAI's own answer — under the coerced `response_format` that is xAI's short-lived URL — rather than a Comfy one, and the rest of the response is unaffected. So one `data` array can mix durable Comfy URLs with expiring partner ones, and a caller that stores or replays this document should not assume every URL in it outlives the call. Such a response IS the answer to the call that produced it, returned and charged exactly as a fully re-hosted one is — but it is replayable only BRIEFLY. A prompt retry carrying the same `Idempotency-Key` — within a few minutes of the original, which is where a connection dropped mid-call puts an SDK's automatic re-send — is answered from the record exactly as any other replay is, because the partner's own link is still alive that soon. A retry after that is answered `409 invalid_input` rather than handed a document whose links may already have expired. The key is consumed either way, so no retry ever re-runs or re-charges; once the replay window has passed, use a new key to run the generation again. A response whose every entry re-hosted cleanly replays normally for the full 24 hours, because a Comfy signed URL is minted with the same lifetime the record is kept for.\n`block_reason` and `usage` are populated WITHOUT any `data` when xAI's input moderation refuses the request, so a caller must key completion off the generated asset rather than off a `200` alone. Key it off an ENTRY of `data` carrying a `url` or a `b64_json`, not off `data` itself: a present-but-asset-less entry is still a billed call that produced nothing, and the container test cannot tell the two apart. That is what the nightly SDK case for this family asserts — the leaves `data[].url` and `data[].b64_json` rather than the `data` container (`testing/e2e/router_sdk/cases.d/xai_image.json`).","example":{"data":[{"mime_type":"image/jpeg","url":"https://example.invalid/xai/grok-imagine-image/generated.jpg"},{"mime_type":"image/jpeg","url":"https://example.invalid/xai/grok-imagine-image/generated-2.jpg"}],"usage":{"cost_in_usd_ticks":200000000}}}}}}}}}},"components":{"schemas":{"XAIGeneratedImage":{"description":"A generated image from xAI","properties":{"b64_json":{"description":"A base64-encoded string representation of the generated image in jpeg encoding (if response_format is b64_json)","type":"string"},"mime_type":{"description":"The MIME type of the generated image (e.g. image/png, image/jpeg, image/webp).","type":"string"},"url":{"description":"A url to the generated image (if response_format is url)","type":"string"}},"type":"object"},"XAIImageGenerationResponse":{"description":"Response from xAI image generation or editing","properties":{"block_reason":{"description":"If the request was blocked by input moderation, contains the block reason","type":"string"},"data":{"description":"A list of generated image objects","items":{"$ref":"#/components/schemas/XAIGeneratedImage"},"type":"array"},"usage":{"$ref":"#/components/schemas/XAIImageUsage"}},"type":"object"},"XAIImageUsage":{"description":"Usage information for the image generation request","properties":{"cost_in_usd_ticks":{"description":"Accurate cost of this request in USD ticks (10,000,000,000 ticks = 1 USD)","type":"integer"}},"type":"object"}},"securitySchemes":{"ApiKeyAuth":{"type":"apiKey","in":"header","name":"X-API-Key"},"BearerAuth":{"type":"http","scheme":"bearer","bearerFormat":"JWT"}}},"x-comfy-router-model-id":"xai/grok-imagine-image-2.0","x-comfy-input-schema-authored":true,"x-comfy-output-schema-authored":true,"x-comfy-router-input-modalities":["text"],"x-comfy-router-output-modalities":["image"]}
+{"openapi":"3.0.2","info":{"title":"xai/grok-imagine-image-2.0","description":"The request body Comfy Router accepts for the model \"xai/grok-imagine-image-2.0\", and the response body it returns. The INPUT schema is the same schema the server validates a call against before it reaches the provider, so what is published and what is enforced cannot differ. `x-comfy-router-input-modalities` records the media kinds the model is documented to accept; the request schema is shared across sibling models and may admit media kinds this model ignores. The OUTPUT schema describes the provider's native result document exactly as Router returns it: Router does not validate, narrow or re-envelope the response, so the output schema is descriptive rather than enforced, and Comfy owns no output shape of its own.","version":"a50bbf97594a"},"paths":{"/v2/models/xai/grok-imagine-image-2.0":{"post":{"operationId":"runRouterModel","summary":"Run xai/grok-imagine-image-2.0 synchronously.","security":[{"BearerAuth":[]},{"ApiKeyAuth":[]}],"requestBody":{"required":true,"content":{"application/json":{"schema":{"description":"Request body for xAI Grok Imagine image generation","example":{"n":1,"prompt":"A single red maple leaf on a plain white background."},"properties":{"aspect_ratio":{"default":"auto","description":"Aspect ratio of the generated image. Defaults to auto for automatically selecting the best ratio for the prompt.","enum":["1:1","3:4","4:3","9:16","16:9","2:3","3:2","9:19.5","19.5:9","9:20","20:9","1:2","2:1","auto"],"type":"string"},"model":{"default":"grok-imagine-image","description":"Model to be used. Supported: grok-imagine-image (default), grok-imagine-image-pro, grok-imagine-image-quality, grok-imagine-image-2.0. Deprecated -beta ids are aliased to their GA model.","nullable":true,"type":"string"},"n":{"default":1,"description":"Number of images to be generated","maximum":10,"minimum":1,"type":"integer"},"prompt":{"description":"Prompt for image generation","type":"string"},"quality":{"description":"Quality of the output image. For grok-imagine-image-2.0 this selects the price tier: low, medium, or auto (xAI's default when omitted; auto serves low for generation and medium for edits). Other models ignore it.","enum":["low","medium","high","auto"],"type":"string"},"resolution":{"default":"1k","description":"Resolution of the generated image. Defaults to 1k.","enum":["1k","2k"],"type":"string"},"response_format":{"default":"url","description":"Response format to return the image in. Can be url or b64_json. Comfy coerces this to `url` on the outbound request — on the Comfy Router dispatch (`POST /v2/models/xai/{model}`) and on this `/proxy/` route too — because image results are served as re-hosted URLs either way. A value IN the enum is therefore accepted and ignored; send `url` or `b64_json` or omit the field, and the answer is the same. A value OUTSIDE the enum — excluding the UNSET spellings each surface recognises, described next — is REJECTED rather than coerced on BOTH surfaces, and only the shape of the refusal differs: this `/proxy/` route answers a `400` at ingress, and the `POST /v2/models/xai/{model}` Router dispatch answers the per-field `422` it owes the enum published here, before any partner call is made. The two surfaces differ on what counts as UNSET. On this `/proxy/` route an omitted field, an explicit `null` and `\"\"` all mean unset and are coerced like any accepted value. The Router dispatch enforces the enum exactly as published, so `null` and `\"\"` are non-members there like any other: OMIT the field to leave it unset on that surface.","enum":["url","b64_json"],"type":"string"},"size":{"description":"Size of the image (not supported)","type":"string"},"style":{"description":"Style of the image (not supported)","type":"string"},"user":{"description":"A unique identifier representing your end-user, which can help xAI to monitor and detect abuse","type":"string"}},"required":["prompt"],"type":"object"}}}},"responses":{"200":{"description":"OK - the model's native JSON output, returned unchanged; the schema describes the provider's terminal result document as Router returns it.","content":{"application/json":{"schema":{"allOf":[{"$ref":"#/components/schemas/XAIImageGenerationResponse"}],"description":"Comfy Router output schema for the xAI Grok Imagine image models: the `POST /v1/images/generations` document xAI returns, forwarded unchanged EXCEPT for the re-hosted `data[]` entries described below - every other field is xAI's own. This family is DIRECT-RETURN — `routerresult/classification.go` records `{provider: xai, endpoint: /v1/images/generations}` as `ReturnModeDirect` with no poll route — so the body a caller receives is the finished generation on the original call rather than a task handle.\nThe generated images are in `data`, and on THIS surface an entry is addressed by its `url`. Router coerces the outbound `response_format` to `url` for this family and re-hosts every image it can fetch onto Comfy storage, replacing xAI's URL with a signed Comfy one and clearing `b64_json`, so a request that asked for `b64_json` is answered exactly as a `url` request is. `mime_type` names the encoding. The `/proxy/` image routes coerce `response_format` to `url` the same way, so they do not preserve a `b64_json` request either.\nDurability is per entry, not per response: an image whose re-host fails keeps xAI's own answer — under the coerced `response_format` that is xAI's short-lived URL — rather than a Comfy one, and the rest of the response is unaffected. So one `data` array can mix durable Comfy URLs with expiring partner ones, and a caller that stores or replays this document should not assume every URL in it outlives the call. Such a response IS the answer to the call that produced it, returned and charged exactly as a fully re-hosted one is — but it is replayable only BRIEFLY. A prompt retry carrying the same `Idempotency-Key` — within a few minutes of the original, which is where a connection dropped mid-call puts an SDK's automatic re-send — is answered from the record exactly as any other replay is, because the partner's own link is still alive that soon. A retry after that is answered `409 invalid_input` rather than handed a document whose links may already have expired. The key is consumed either way, so no retry ever re-runs or re-charges; once the replay window has passed, use a new key to run the generation again. A response whose every entry re-hosted cleanly replays normally for the full 24 hours, because a Comfy signed URL is minted with the same lifetime the record is kept for.\n`block_reason` and `usage` are populated WITHOUT any `data` when xAI's input moderation refuses the request, so a caller must key completion off the generated asset rather than off a `200` alone. Key it off an ENTRY of `data` carrying a `url` or a `b64_json`, not off `data` itself: a present-but-asset-less entry is still a billed call that produced nothing, and the container test cannot tell the two apart. That is what the nightly SDK case for this family asserts — the leaves `data[].url` and `data[].b64_json` rather than the `data` container (`testing/e2e/router_sdk/cases.d/xai_image.json`).","example":{"data":[{"mime_type":"image/jpeg","url":"https://example.invalid/xai/grok-imagine-image/generated.jpg"},{"mime_type":"image/jpeg","url":"https://example.invalid/xai/grok-imagine-image/generated-2.jpg"}],"usage":{"cost_in_usd_ticks":200000000}}}}}}}}}},"components":{"schemas":{"XAIGeneratedImage":{"description":"A generated image from xAI","properties":{"b64_json":{"description":"A base64-encoded string representation of the generated image in jpeg encoding (if response_format is b64_json)","type":"string"},"mime_type":{"description":"The MIME type of the generated image (e.g. image/png, image/jpeg, image/webp).","type":"string"},"url":{"description":"A url to the generated image (if response_format is url)","type":"string"}},"type":"object"},"XAIImageGenerationResponse":{"description":"Response from xAI image generation or editing","properties":{"block_reason":{"description":"If the request was blocked by input moderation, contains the block reason","type":"string"},"data":{"description":"A list of generated image objects","items":{"$ref":"#/components/schemas/XAIGeneratedImage"},"type":"array"},"usage":{"$ref":"#/components/schemas/XAIImageUsage"}},"type":"object"},"XAIImageUsage":{"description":"Usage information for the image generation request","properties":{"cost_in_usd_ticks":{"description":"Accurate cost of this request in USD ticks (10,000,000,000 ticks = 1 USD)","type":"integer"}},"type":"object"}},"securitySchemes":{"ApiKeyAuth":{"type":"apiKey","in":"header","name":"X-API-Key"},"BearerAuth":{"type":"http","scheme":"bearer","bearerFormat":"JWT"}}},"x-comfy-router-model-id":"xai/grok-imagine-image-2.0","x-comfy-input-schema-authored":true,"x-comfy-output-schema-authored":true,"x-comfy-router-input-modalities":["text"],"x-comfy-router-output-modalities":["image"]}
diff --git a/router-schemas/xai/grok-imagine-image-pro.json b/router-schemas/xai/grok-imagine-image-pro.json
index 64636f99ee..839aae102b 100644
--- a/router-schemas/xai/grok-imagine-image-pro.json
+++ b/router-schemas/xai/grok-imagine-image-pro.json
@@ -1 +1 @@
-{"openapi":"3.0.2","info":{"title":"xai/grok-imagine-image-pro","description":"The request body Comfy Router accepts for the model \"xai/grok-imagine-image-pro\", and the response body it returns. The INPUT schema is the same schema the server validates a call against before it reaches the provider, so what is published and what is enforced cannot differ. `x-comfy-router-input-modalities` records the media kinds the model is documented to accept; the request schema is shared across sibling models and may admit media kinds this model ignores. The OUTPUT schema describes the provider's native result document exactly as Router returns it: Router does not validate, narrow or re-envelope the response, so the output schema is descriptive rather than enforced, and Comfy owns no output shape of its own.","version":"755fafabd68d"},"paths":{"/v2/models/xai/grok-imagine-image-pro":{"post":{"operationId":"runRouterModel","summary":"Run xai/grok-imagine-image-pro synchronously.","security":[{"BearerAuth":[]},{"ApiKeyAuth":[]}],"requestBody":{"required":true,"content":{"application/json":{"schema":{"description":"Request body for xAI Grok Imagine image generation","example":{"n":1,"prompt":"A single red maple leaf on a plain white background."},"properties":{"aspect_ratio":{"default":"auto","description":"Aspect ratio of the generated image. Defaults to auto for automatically selecting the best ratio for the prompt.","enum":["1:1","3:4","4:3","9:16","16:9","2:3","3:2","9:19.5","19.5:9","9:20","20:9","1:2","2:1","auto"],"type":"string"},"model":{"default":"grok-imagine-image","description":"Model to be used. Supported: grok-imagine-image (default), grok-imagine-image-pro, grok-imagine-image-quality, grok-imagine-image-2.0. Deprecated -beta ids are aliased to their GA model.","nullable":true,"type":"string"},"n":{"default":1,"description":"Number of images to be generated","maximum":10,"minimum":1,"type":"integer"},"prompt":{"description":"Prompt for image generation","type":"string"},"quality":{"description":"Quality of the output image. For grok-imagine-image-2.0 this selects the price tier (low/medium; medium is the default); other models currently ignore it.","enum":["low","medium","high"],"type":"string"},"resolution":{"default":"1k","description":"Resolution of the generated image. Defaults to 1k.","enum":["1k","2k"],"type":"string"},"response_format":{"default":"url","description":"Response format to return the image in. Can be url or b64_json. Comfy coerces this to `url` on the outbound request — on the Comfy Router dispatch (`POST /v2/models/xai/{model}`) and on this `/proxy/` route too — because image results are served as re-hosted URLs either way. A value IN the enum is therefore accepted and ignored; send `url` or `b64_json` or omit the field, and the answer is the same. A value OUTSIDE the enum — excluding the UNSET spellings each surface recognises, described next — is REJECTED rather than coerced on BOTH surfaces, and only the shape of the refusal differs: this `/proxy/` route answers a `400` at ingress, and the `POST /v2/models/xai/{model}` Router dispatch answers the per-field `422` it owes the enum published here, before any partner call is made. The two surfaces differ on what counts as UNSET. On this `/proxy/` route an omitted field, an explicit `null` and `\"\"` all mean unset and are coerced like any accepted value. The Router dispatch enforces the enum exactly as published, so `null` and `\"\"` are non-members there like any other: OMIT the field to leave it unset on that surface.","enum":["url","b64_json"],"type":"string"},"size":{"description":"Size of the image (not supported)","type":"string"},"style":{"description":"Style of the image (not supported)","type":"string"},"user":{"description":"A unique identifier representing your end-user, which can help xAI to monitor and detect abuse","type":"string"}},"required":["prompt"],"type":"object"}}}},"responses":{"200":{"description":"OK - the model's native JSON output, returned unchanged; the schema describes the provider's terminal result document as Router returns it.","content":{"application/json":{"schema":{"allOf":[{"$ref":"#/components/schemas/XAIImageGenerationResponse"}],"description":"Comfy Router output schema for the xAI Grok Imagine image models: the `POST /v1/images/generations` document xAI returns, forwarded unchanged EXCEPT for the re-hosted `data[]` entries described below - every other field is xAI's own. This family is DIRECT-RETURN — `routerresult/classification.go` records `{provider: xai, endpoint: /v1/images/generations}` as `ReturnModeDirect` with no poll route — so the body a caller receives is the finished generation on the original call rather than a task handle.\nThe generated images are in `data`, and on THIS surface an entry is addressed by its `url`. Router coerces the outbound `response_format` to `url` for this family and re-hosts every image it can fetch onto Comfy storage, replacing xAI's URL with a signed Comfy one and clearing `b64_json`, so a request that asked for `b64_json` is answered exactly as a `url` request is. `mime_type` names the encoding. The `/proxy/` image routes coerce `response_format` to `url` the same way, so they do not preserve a `b64_json` request either.\nDurability is per entry, not per response: an image whose re-host fails keeps xAI's own answer — under the coerced `response_format` that is xAI's short-lived URL — rather than a Comfy one, and the rest of the response is unaffected. So one `data` array can mix durable Comfy URLs with expiring partner ones, and a caller that stores or replays this document should not assume every URL in it outlives the call. Such a response IS the answer to the call that produced it, returned and charged exactly as a fully re-hosted one is — but it is replayable only BRIEFLY. A prompt retry carrying the same `Idempotency-Key` — within a few minutes of the original, which is where a connection dropped mid-call puts an SDK's automatic re-send — is answered from the record exactly as any other replay is, because the partner's own link is still alive that soon. A retry after that is answered `409 invalid_input` rather than handed a document whose links may already have expired. The key is consumed either way, so no retry ever re-runs or re-charges; once the replay window has passed, use a new key to run the generation again. A response whose every entry re-hosted cleanly replays normally for the full 24 hours, because a Comfy signed URL is minted with the same lifetime the record is kept for.\n`block_reason` and `usage` are populated WITHOUT any `data` when xAI's input moderation refuses the request, so a caller must key completion off the generated asset rather than off a `200` alone. Key it off an ENTRY of `data` carrying a `url` or a `b64_json`, not off `data` itself: a present-but-asset-less entry is still a billed call that produced nothing, and the container test cannot tell the two apart. That is what the nightly SDK case for this family asserts — the leaves `data[].url` and `data[].b64_json` rather than the `data` container (`testing/e2e/router_sdk/cases.d/xai_image.json`).","example":{"data":[{"mime_type":"image/jpeg","url":"https://example.invalid/xai/grok-imagine-image/generated.jpg"},{"mime_type":"image/jpeg","url":"https://example.invalid/xai/grok-imagine-image/generated-2.jpg"}],"usage":{"cost_in_usd_ticks":200000000}}}}}}}}}},"components":{"schemas":{"XAIGeneratedImage":{"description":"A generated image from xAI","properties":{"b64_json":{"description":"A base64-encoded string representation of the generated image in jpeg encoding (if response_format is b64_json)","type":"string"},"mime_type":{"description":"The MIME type of the generated image (e.g. image/png, image/jpeg, image/webp).","type":"string"},"url":{"description":"A url to the generated image (if response_format is url)","type":"string"}},"type":"object"},"XAIImageGenerationResponse":{"description":"Response from xAI image generation or editing","properties":{"block_reason":{"description":"If the request was blocked by input moderation, contains the block reason","type":"string"},"data":{"description":"A list of generated image objects","items":{"$ref":"#/components/schemas/XAIGeneratedImage"},"type":"array"},"usage":{"$ref":"#/components/schemas/XAIImageUsage"}},"type":"object"},"XAIImageUsage":{"description":"Usage information for the image generation request","properties":{"cost_in_usd_ticks":{"description":"Accurate cost of this request in USD ticks (10,000,000,000 ticks = 1 USD)","type":"integer"}},"type":"object"}},"securitySchemes":{"ApiKeyAuth":{"type":"apiKey","in":"header","name":"X-API-Key"},"BearerAuth":{"type":"http","scheme":"bearer","bearerFormat":"JWT"}}},"x-comfy-router-model-id":"xai/grok-imagine-image-pro","x-comfy-input-schema-authored":true,"x-comfy-output-schema-authored":true,"x-comfy-router-input-modalities":["text"],"x-comfy-router-output-modalities":["image"]}
+{"openapi":"3.0.2","info":{"title":"xai/grok-imagine-image-pro","description":"The request body Comfy Router accepts for the model \"xai/grok-imagine-image-pro\", and the response body it returns. The INPUT schema is the same schema the server validates a call against before it reaches the provider, so what is published and what is enforced cannot differ. `x-comfy-router-input-modalities` records the media kinds the model is documented to accept; the request schema is shared across sibling models and may admit media kinds this model ignores. The OUTPUT schema describes the provider's native result document exactly as Router returns it: Router does not validate, narrow or re-envelope the response, so the output schema is descriptive rather than enforced, and Comfy owns no output shape of its own.","version":"a50bbf97594a"},"paths":{"/v2/models/xai/grok-imagine-image-pro":{"post":{"operationId":"runRouterModel","summary":"Run xai/grok-imagine-image-pro synchronously.","security":[{"BearerAuth":[]},{"ApiKeyAuth":[]}],"requestBody":{"required":true,"content":{"application/json":{"schema":{"description":"Request body for xAI Grok Imagine image generation","example":{"n":1,"prompt":"A single red maple leaf on a plain white background."},"properties":{"aspect_ratio":{"default":"auto","description":"Aspect ratio of the generated image. Defaults to auto for automatically selecting the best ratio for the prompt.","enum":["1:1","3:4","4:3","9:16","16:9","2:3","3:2","9:19.5","19.5:9","9:20","20:9","1:2","2:1","auto"],"type":"string"},"model":{"default":"grok-imagine-image","description":"Model to be used. Supported: grok-imagine-image (default), grok-imagine-image-pro, grok-imagine-image-quality, grok-imagine-image-2.0. Deprecated -beta ids are aliased to their GA model.","nullable":true,"type":"string"},"n":{"default":1,"description":"Number of images to be generated","maximum":10,"minimum":1,"type":"integer"},"prompt":{"description":"Prompt for image generation","type":"string"},"quality":{"description":"Quality of the output image. For grok-imagine-image-2.0 this selects the price tier: low, medium, or auto (xAI's default when omitted; auto serves low for generation and medium for edits). Other models ignore it.","enum":["low","medium","high","auto"],"type":"string"},"resolution":{"default":"1k","description":"Resolution of the generated image. Defaults to 1k.","enum":["1k","2k"],"type":"string"},"response_format":{"default":"url","description":"Response format to return the image in. Can be url or b64_json. Comfy coerces this to `url` on the outbound request — on the Comfy Router dispatch (`POST /v2/models/xai/{model}`) and on this `/proxy/` route too — because image results are served as re-hosted URLs either way. A value IN the enum is therefore accepted and ignored; send `url` or `b64_json` or omit the field, and the answer is the same. A value OUTSIDE the enum — excluding the UNSET spellings each surface recognises, described next — is REJECTED rather than coerced on BOTH surfaces, and only the shape of the refusal differs: this `/proxy/` route answers a `400` at ingress, and the `POST /v2/models/xai/{model}` Router dispatch answers the per-field `422` it owes the enum published here, before any partner call is made. The two surfaces differ on what counts as UNSET. On this `/proxy/` route an omitted field, an explicit `null` and `\"\"` all mean unset and are coerced like any accepted value. The Router dispatch enforces the enum exactly as published, so `null` and `\"\"` are non-members there like any other: OMIT the field to leave it unset on that surface.","enum":["url","b64_json"],"type":"string"},"size":{"description":"Size of the image (not supported)","type":"string"},"style":{"description":"Style of the image (not supported)","type":"string"},"user":{"description":"A unique identifier representing your end-user, which can help xAI to monitor and detect abuse","type":"string"}},"required":["prompt"],"type":"object"}}}},"responses":{"200":{"description":"OK - the model's native JSON output, returned unchanged; the schema describes the provider's terminal result document as Router returns it.","content":{"application/json":{"schema":{"allOf":[{"$ref":"#/components/schemas/XAIImageGenerationResponse"}],"description":"Comfy Router output schema for the xAI Grok Imagine image models: the `POST /v1/images/generations` document xAI returns, forwarded unchanged EXCEPT for the re-hosted `data[]` entries described below - every other field is xAI's own. This family is DIRECT-RETURN — `routerresult/classification.go` records `{provider: xai, endpoint: /v1/images/generations}` as `ReturnModeDirect` with no poll route — so the body a caller receives is the finished generation on the original call rather than a task handle.\nThe generated images are in `data`, and on THIS surface an entry is addressed by its `url`. Router coerces the outbound `response_format` to `url` for this family and re-hosts every image it can fetch onto Comfy storage, replacing xAI's URL with a signed Comfy one and clearing `b64_json`, so a request that asked for `b64_json` is answered exactly as a `url` request is. `mime_type` names the encoding. The `/proxy/` image routes coerce `response_format` to `url` the same way, so they do not preserve a `b64_json` request either.\nDurability is per entry, not per response: an image whose re-host fails keeps xAI's own answer — under the coerced `response_format` that is xAI's short-lived URL — rather than a Comfy one, and the rest of the response is unaffected. So one `data` array can mix durable Comfy URLs with expiring partner ones, and a caller that stores or replays this document should not assume every URL in it outlives the call. Such a response IS the answer to the call that produced it, returned and charged exactly as a fully re-hosted one is — but it is replayable only BRIEFLY. A prompt retry carrying the same `Idempotency-Key` — within a few minutes of the original, which is where a connection dropped mid-call puts an SDK's automatic re-send — is answered from the record exactly as any other replay is, because the partner's own link is still alive that soon. A retry after that is answered `409 invalid_input` rather than handed a document whose links may already have expired. The key is consumed either way, so no retry ever re-runs or re-charges; once the replay window has passed, use a new key to run the generation again. A response whose every entry re-hosted cleanly replays normally for the full 24 hours, because a Comfy signed URL is minted with the same lifetime the record is kept for.\n`block_reason` and `usage` are populated WITHOUT any `data` when xAI's input moderation refuses the request, so a caller must key completion off the generated asset rather than off a `200` alone. Key it off an ENTRY of `data` carrying a `url` or a `b64_json`, not off `data` itself: a present-but-asset-less entry is still a billed call that produced nothing, and the container test cannot tell the two apart. That is what the nightly SDK case for this family asserts — the leaves `data[].url` and `data[].b64_json` rather than the `data` container (`testing/e2e/router_sdk/cases.d/xai_image.json`).","example":{"data":[{"mime_type":"image/jpeg","url":"https://example.invalid/xai/grok-imagine-image/generated.jpg"},{"mime_type":"image/jpeg","url":"https://example.invalid/xai/grok-imagine-image/generated-2.jpg"}],"usage":{"cost_in_usd_ticks":200000000}}}}}}}}}},"components":{"schemas":{"XAIGeneratedImage":{"description":"A generated image from xAI","properties":{"b64_json":{"description":"A base64-encoded string representation of the generated image in jpeg encoding (if response_format is b64_json)","type":"string"},"mime_type":{"description":"The MIME type of the generated image (e.g. image/png, image/jpeg, image/webp).","type":"string"},"url":{"description":"A url to the generated image (if response_format is url)","type":"string"}},"type":"object"},"XAIImageGenerationResponse":{"description":"Response from xAI image generation or editing","properties":{"block_reason":{"description":"If the request was blocked by input moderation, contains the block reason","type":"string"},"data":{"description":"A list of generated image objects","items":{"$ref":"#/components/schemas/XAIGeneratedImage"},"type":"array"},"usage":{"$ref":"#/components/schemas/XAIImageUsage"}},"type":"object"},"XAIImageUsage":{"description":"Usage information for the image generation request","properties":{"cost_in_usd_ticks":{"description":"Accurate cost of this request in USD ticks (10,000,000,000 ticks = 1 USD)","type":"integer"}},"type":"object"}},"securitySchemes":{"ApiKeyAuth":{"type":"apiKey","in":"header","name":"X-API-Key"},"BearerAuth":{"type":"http","scheme":"bearer","bearerFormat":"JWT"}}},"x-comfy-router-model-id":"xai/grok-imagine-image-pro","x-comfy-input-schema-authored":true,"x-comfy-output-schema-authored":true,"x-comfy-router-input-modalities":["text"],"x-comfy-router-output-modalities":["image"]}
diff --git a/router-schemas/xai/grok-imagine-image-quality.json b/router-schemas/xai/grok-imagine-image-quality.json
index b477466b42..a5374c4907 100644
--- a/router-schemas/xai/grok-imagine-image-quality.json
+++ b/router-schemas/xai/grok-imagine-image-quality.json
@@ -1 +1 @@
-{"openapi":"3.0.2","info":{"title":"xai/grok-imagine-image-quality","description":"The request body Comfy Router accepts for the model \"xai/grok-imagine-image-quality\", and the response body it returns. The INPUT schema is the same schema the server validates a call against before it reaches the provider, so what is published and what is enforced cannot differ. `x-comfy-router-input-modalities` records the media kinds the model is documented to accept; the request schema is shared across sibling models and may admit media kinds this model ignores. The OUTPUT schema describes the provider's native result document exactly as Router returns it: Router does not validate, narrow or re-envelope the response, so the output schema is descriptive rather than enforced, and Comfy owns no output shape of its own.","version":"755fafabd68d"},"paths":{"/v2/models/xai/grok-imagine-image-quality":{"post":{"operationId":"runRouterModel","summary":"Run xai/grok-imagine-image-quality synchronously.","security":[{"BearerAuth":[]},{"ApiKeyAuth":[]}],"requestBody":{"required":true,"content":{"application/json":{"schema":{"description":"Request body for xAI Grok Imagine image generation","example":{"n":1,"prompt":"A single red maple leaf on a plain white background."},"properties":{"aspect_ratio":{"default":"auto","description":"Aspect ratio of the generated image. Defaults to auto for automatically selecting the best ratio for the prompt.","enum":["1:1","3:4","4:3","9:16","16:9","2:3","3:2","9:19.5","19.5:9","9:20","20:9","1:2","2:1","auto"],"type":"string"},"model":{"default":"grok-imagine-image","description":"Model to be used. Supported: grok-imagine-image (default), grok-imagine-image-pro, grok-imagine-image-quality, grok-imagine-image-2.0. Deprecated -beta ids are aliased to their GA model.","nullable":true,"type":"string"},"n":{"default":1,"description":"Number of images to be generated","maximum":10,"minimum":1,"type":"integer"},"prompt":{"description":"Prompt for image generation","type":"string"},"quality":{"description":"Quality of the output image. For grok-imagine-image-2.0 this selects the price tier (low/medium; medium is the default); other models currently ignore it.","enum":["low","medium","high"],"type":"string"},"resolution":{"default":"1k","description":"Resolution of the generated image. Defaults to 1k.","enum":["1k","2k"],"type":"string"},"response_format":{"default":"url","description":"Response format to return the image in. Can be url or b64_json. Comfy coerces this to `url` on the outbound request — on the Comfy Router dispatch (`POST /v2/models/xai/{model}`) and on this `/proxy/` route too — because image results are served as re-hosted URLs either way. A value IN the enum is therefore accepted and ignored; send `url` or `b64_json` or omit the field, and the answer is the same. A value OUTSIDE the enum — excluding the UNSET spellings each surface recognises, described next — is REJECTED rather than coerced on BOTH surfaces, and only the shape of the refusal differs: this `/proxy/` route answers a `400` at ingress, and the `POST /v2/models/xai/{model}` Router dispatch answers the per-field `422` it owes the enum published here, before any partner call is made. The two surfaces differ on what counts as UNSET. On this `/proxy/` route an omitted field, an explicit `null` and `\"\"` all mean unset and are coerced like any accepted value. The Router dispatch enforces the enum exactly as published, so `null` and `\"\"` are non-members there like any other: OMIT the field to leave it unset on that surface.","enum":["url","b64_json"],"type":"string"},"size":{"description":"Size of the image (not supported)","type":"string"},"style":{"description":"Style of the image (not supported)","type":"string"},"user":{"description":"A unique identifier representing your end-user, which can help xAI to monitor and detect abuse","type":"string"}},"required":["prompt"],"type":"object"}}}},"responses":{"200":{"description":"OK - the model's native JSON output, returned unchanged; the schema describes the provider's terminal result document as Router returns it.","content":{"application/json":{"schema":{"allOf":[{"$ref":"#/components/schemas/XAIImageGenerationResponse"}],"description":"Comfy Router output schema for the xAI Grok Imagine image models: the `POST /v1/images/generations` document xAI returns, forwarded unchanged EXCEPT for the re-hosted `data[]` entries described below - every other field is xAI's own. This family is DIRECT-RETURN — `routerresult/classification.go` records `{provider: xai, endpoint: /v1/images/generations}` as `ReturnModeDirect` with no poll route — so the body a caller receives is the finished generation on the original call rather than a task handle.\nThe generated images are in `data`, and on THIS surface an entry is addressed by its `url`. Router coerces the outbound `response_format` to `url` for this family and re-hosts every image it can fetch onto Comfy storage, replacing xAI's URL with a signed Comfy one and clearing `b64_json`, so a request that asked for `b64_json` is answered exactly as a `url` request is. `mime_type` names the encoding. The `/proxy/` image routes coerce `response_format` to `url` the same way, so they do not preserve a `b64_json` request either.\nDurability is per entry, not per response: an image whose re-host fails keeps xAI's own answer — under the coerced `response_format` that is xAI's short-lived URL — rather than a Comfy one, and the rest of the response is unaffected. So one `data` array can mix durable Comfy URLs with expiring partner ones, and a caller that stores or replays this document should not assume every URL in it outlives the call. Such a response IS the answer to the call that produced it, returned and charged exactly as a fully re-hosted one is — but it is replayable only BRIEFLY. A prompt retry carrying the same `Idempotency-Key` — within a few minutes of the original, which is where a connection dropped mid-call puts an SDK's automatic re-send — is answered from the record exactly as any other replay is, because the partner's own link is still alive that soon. A retry after that is answered `409 invalid_input` rather than handed a document whose links may already have expired. The key is consumed either way, so no retry ever re-runs or re-charges; once the replay window has passed, use a new key to run the generation again. A response whose every entry re-hosted cleanly replays normally for the full 24 hours, because a Comfy signed URL is minted with the same lifetime the record is kept for.\n`block_reason` and `usage` are populated WITHOUT any `data` when xAI's input moderation refuses the request, so a caller must key completion off the generated asset rather than off a `200` alone. Key it off an ENTRY of `data` carrying a `url` or a `b64_json`, not off `data` itself: a present-but-asset-less entry is still a billed call that produced nothing, and the container test cannot tell the two apart. That is what the nightly SDK case for this family asserts — the leaves `data[].url` and `data[].b64_json` rather than the `data` container (`testing/e2e/router_sdk/cases.d/xai_image.json`).","example":{"data":[{"mime_type":"image/jpeg","url":"https://example.invalid/xai/grok-imagine-image/generated.jpg"},{"mime_type":"image/jpeg","url":"https://example.invalid/xai/grok-imagine-image/generated-2.jpg"}],"usage":{"cost_in_usd_ticks":200000000}}}}}}}}}},"components":{"schemas":{"XAIGeneratedImage":{"description":"A generated image from xAI","properties":{"b64_json":{"description":"A base64-encoded string representation of the generated image in jpeg encoding (if response_format is b64_json)","type":"string"},"mime_type":{"description":"The MIME type of the generated image (e.g. image/png, image/jpeg, image/webp).","type":"string"},"url":{"description":"A url to the generated image (if response_format is url)","type":"string"}},"type":"object"},"XAIImageGenerationResponse":{"description":"Response from xAI image generation or editing","properties":{"block_reason":{"description":"If the request was blocked by input moderation, contains the block reason","type":"string"},"data":{"description":"A list of generated image objects","items":{"$ref":"#/components/schemas/XAIGeneratedImage"},"type":"array"},"usage":{"$ref":"#/components/schemas/XAIImageUsage"}},"type":"object"},"XAIImageUsage":{"description":"Usage information for the image generation request","properties":{"cost_in_usd_ticks":{"description":"Accurate cost of this request in USD ticks (10,000,000,000 ticks = 1 USD)","type":"integer"}},"type":"object"}},"securitySchemes":{"ApiKeyAuth":{"type":"apiKey","in":"header","name":"X-API-Key"},"BearerAuth":{"type":"http","scheme":"bearer","bearerFormat":"JWT"}}},"x-comfy-router-model-id":"xai/grok-imagine-image-quality","x-comfy-input-schema-authored":true,"x-comfy-output-schema-authored":true,"x-comfy-router-input-modalities":["text"],"x-comfy-router-output-modalities":["image"]}
+{"openapi":"3.0.2","info":{"title":"xai/grok-imagine-image-quality","description":"The request body Comfy Router accepts for the model \"xai/grok-imagine-image-quality\", and the response body it returns. The INPUT schema is the same schema the server validates a call against before it reaches the provider, so what is published and what is enforced cannot differ. `x-comfy-router-input-modalities` records the media kinds the model is documented to accept; the request schema is shared across sibling models and may admit media kinds this model ignores. The OUTPUT schema describes the provider's native result document exactly as Router returns it: Router does not validate, narrow or re-envelope the response, so the output schema is descriptive rather than enforced, and Comfy owns no output shape of its own.","version":"a50bbf97594a"},"paths":{"/v2/models/xai/grok-imagine-image-quality":{"post":{"operationId":"runRouterModel","summary":"Run xai/grok-imagine-image-quality synchronously.","security":[{"BearerAuth":[]},{"ApiKeyAuth":[]}],"requestBody":{"required":true,"content":{"application/json":{"schema":{"description":"Request body for xAI Grok Imagine image generation","example":{"n":1,"prompt":"A single red maple leaf on a plain white background."},"properties":{"aspect_ratio":{"default":"auto","description":"Aspect ratio of the generated image. Defaults to auto for automatically selecting the best ratio for the prompt.","enum":["1:1","3:4","4:3","9:16","16:9","2:3","3:2","9:19.5","19.5:9","9:20","20:9","1:2","2:1","auto"],"type":"string"},"model":{"default":"grok-imagine-image","description":"Model to be used. Supported: grok-imagine-image (default), grok-imagine-image-pro, grok-imagine-image-quality, grok-imagine-image-2.0. Deprecated -beta ids are aliased to their GA model.","nullable":true,"type":"string"},"n":{"default":1,"description":"Number of images to be generated","maximum":10,"minimum":1,"type":"integer"},"prompt":{"description":"Prompt for image generation","type":"string"},"quality":{"description":"Quality of the output image. For grok-imagine-image-2.0 this selects the price tier: low, medium, or auto (xAI's default when omitted; auto serves low for generation and medium for edits). Other models ignore it.","enum":["low","medium","high","auto"],"type":"string"},"resolution":{"default":"1k","description":"Resolution of the generated image. Defaults to 1k.","enum":["1k","2k"],"type":"string"},"response_format":{"default":"url","description":"Response format to return the image in. Can be url or b64_json. Comfy coerces this to `url` on the outbound request — on the Comfy Router dispatch (`POST /v2/models/xai/{model}`) and on this `/proxy/` route too — because image results are served as re-hosted URLs either way. A value IN the enum is therefore accepted and ignored; send `url` or `b64_json` or omit the field, and the answer is the same. A value OUTSIDE the enum — excluding the UNSET spellings each surface recognises, described next — is REJECTED rather than coerced on BOTH surfaces, and only the shape of the refusal differs: this `/proxy/` route answers a `400` at ingress, and the `POST /v2/models/xai/{model}` Router dispatch answers the per-field `422` it owes the enum published here, before any partner call is made. The two surfaces differ on what counts as UNSET. On this `/proxy/` route an omitted field, an explicit `null` and `\"\"` all mean unset and are coerced like any accepted value. The Router dispatch enforces the enum exactly as published, so `null` and `\"\"` are non-members there like any other: OMIT the field to leave it unset on that surface.","enum":["url","b64_json"],"type":"string"},"size":{"description":"Size of the image (not supported)","type":"string"},"style":{"description":"Style of the image (not supported)","type":"string"},"user":{"description":"A unique identifier representing your end-user, which can help xAI to monitor and detect abuse","type":"string"}},"required":["prompt"],"type":"object"}}}},"responses":{"200":{"description":"OK - the model's native JSON output, returned unchanged; the schema describes the provider's terminal result document as Router returns it.","content":{"application/json":{"schema":{"allOf":[{"$ref":"#/components/schemas/XAIImageGenerationResponse"}],"description":"Comfy Router output schema for the xAI Grok Imagine image models: the `POST /v1/images/generations` document xAI returns, forwarded unchanged EXCEPT for the re-hosted `data[]` entries described below - every other field is xAI's own. This family is DIRECT-RETURN — `routerresult/classification.go` records `{provider: xai, endpoint: /v1/images/generations}` as `ReturnModeDirect` with no poll route — so the body a caller receives is the finished generation on the original call rather than a task handle.\nThe generated images are in `data`, and on THIS surface an entry is addressed by its `url`. Router coerces the outbound `response_format` to `url` for this family and re-hosts every image it can fetch onto Comfy storage, replacing xAI's URL with a signed Comfy one and clearing `b64_json`, so a request that asked for `b64_json` is answered exactly as a `url` request is. `mime_type` names the encoding. The `/proxy/` image routes coerce `response_format` to `url` the same way, so they do not preserve a `b64_json` request either.\nDurability is per entry, not per response: an image whose re-host fails keeps xAI's own answer — under the coerced `response_format` that is xAI's short-lived URL — rather than a Comfy one, and the rest of the response is unaffected. So one `data` array can mix durable Comfy URLs with expiring partner ones, and a caller that stores or replays this document should not assume every URL in it outlives the call. Such a response IS the answer to the call that produced it, returned and charged exactly as a fully re-hosted one is — but it is replayable only BRIEFLY. A prompt retry carrying the same `Idempotency-Key` — within a few minutes of the original, which is where a connection dropped mid-call puts an SDK's automatic re-send — is answered from the record exactly as any other replay is, because the partner's own link is still alive that soon. A retry after that is answered `409 invalid_input` rather than handed a document whose links may already have expired. The key is consumed either way, so no retry ever re-runs or re-charges; once the replay window has passed, use a new key to run the generation again. A response whose every entry re-hosted cleanly replays normally for the full 24 hours, because a Comfy signed URL is minted with the same lifetime the record is kept for.\n`block_reason` and `usage` are populated WITHOUT any `data` when xAI's input moderation refuses the request, so a caller must key completion off the generated asset rather than off a `200` alone. Key it off an ENTRY of `data` carrying a `url` or a `b64_json`, not off `data` itself: a present-but-asset-less entry is still a billed call that produced nothing, and the container test cannot tell the two apart. That is what the nightly SDK case for this family asserts — the leaves `data[].url` and `data[].b64_json` rather than the `data` container (`testing/e2e/router_sdk/cases.d/xai_image.json`).","example":{"data":[{"mime_type":"image/jpeg","url":"https://example.invalid/xai/grok-imagine-image/generated.jpg"},{"mime_type":"image/jpeg","url":"https://example.invalid/xai/grok-imagine-image/generated-2.jpg"}],"usage":{"cost_in_usd_ticks":200000000}}}}}}}}}},"components":{"schemas":{"XAIGeneratedImage":{"description":"A generated image from xAI","properties":{"b64_json":{"description":"A base64-encoded string representation of the generated image in jpeg encoding (if response_format is b64_json)","type":"string"},"mime_type":{"description":"The MIME type of the generated image (e.g. image/png, image/jpeg, image/webp).","type":"string"},"url":{"description":"A url to the generated image (if response_format is url)","type":"string"}},"type":"object"},"XAIImageGenerationResponse":{"description":"Response from xAI image generation or editing","properties":{"block_reason":{"description":"If the request was blocked by input moderation, contains the block reason","type":"string"},"data":{"description":"A list of generated image objects","items":{"$ref":"#/components/schemas/XAIGeneratedImage"},"type":"array"},"usage":{"$ref":"#/components/schemas/XAIImageUsage"}},"type":"object"},"XAIImageUsage":{"description":"Usage information for the image generation request","properties":{"cost_in_usd_ticks":{"description":"Accurate cost of this request in USD ticks (10,000,000,000 ticks = 1 USD)","type":"integer"}},"type":"object"}},"securitySchemes":{"ApiKeyAuth":{"type":"apiKey","in":"header","name":"X-API-Key"},"BearerAuth":{"type":"http","scheme":"bearer","bearerFormat":"JWT"}}},"x-comfy-router-model-id":"xai/grok-imagine-image-quality","x-comfy-input-schema-authored":true,"x-comfy-output-schema-authored":true,"x-comfy-router-input-modalities":["text"],"x-comfy-router-output-modalities":["image"]}
diff --git a/router-schemas/xai/grok-imagine-image.json b/router-schemas/xai/grok-imagine-image.json
index 3cf8dfff08..84b4e0dddb 100644
--- a/router-schemas/xai/grok-imagine-image.json
+++ b/router-schemas/xai/grok-imagine-image.json
@@ -1 +1 @@
-{"openapi":"3.0.2","info":{"title":"xai/grok-imagine-image","description":"The request body Comfy Router accepts for the model \"xai/grok-imagine-image\", and the response body it returns. The INPUT schema is the same schema the server validates a call against before it reaches the provider, so what is published and what is enforced cannot differ. `x-comfy-router-input-modalities` records the media kinds the model is documented to accept; the request schema is shared across sibling models and may admit media kinds this model ignores. The OUTPUT schema describes the provider's native result document exactly as Router returns it: Router does not validate, narrow or re-envelope the response, so the output schema is descriptive rather than enforced, and Comfy owns no output shape of its own.","version":"755fafabd68d"},"paths":{"/v2/models/xai/grok-imagine-image":{"post":{"operationId":"runRouterModel","summary":"Run xai/grok-imagine-image synchronously.","security":[{"BearerAuth":[]},{"ApiKeyAuth":[]}],"requestBody":{"required":true,"content":{"application/json":{"schema":{"description":"Request body for xAI Grok Imagine image generation","example":{"n":1,"prompt":"A single red maple leaf on a plain white background."},"properties":{"aspect_ratio":{"default":"auto","description":"Aspect ratio of the generated image. Defaults to auto for automatically selecting the best ratio for the prompt.","enum":["1:1","3:4","4:3","9:16","16:9","2:3","3:2","9:19.5","19.5:9","9:20","20:9","1:2","2:1","auto"],"type":"string"},"model":{"default":"grok-imagine-image","description":"Model to be used. Supported: grok-imagine-image (default), grok-imagine-image-pro, grok-imagine-image-quality, grok-imagine-image-2.0. Deprecated -beta ids are aliased to their GA model.","nullable":true,"type":"string"},"n":{"default":1,"description":"Number of images to be generated","maximum":10,"minimum":1,"type":"integer"},"prompt":{"description":"Prompt for image generation","type":"string"},"quality":{"description":"Quality of the output image. For grok-imagine-image-2.0 this selects the price tier (low/medium; medium is the default); other models currently ignore it.","enum":["low","medium","high"],"type":"string"},"resolution":{"default":"1k","description":"Resolution of the generated image. Defaults to 1k.","enum":["1k","2k"],"type":"string"},"response_format":{"default":"url","description":"Response format to return the image in. Can be url or b64_json. Comfy coerces this to `url` on the outbound request — on the Comfy Router dispatch (`POST /v2/models/xai/{model}`) and on this `/proxy/` route too — because image results are served as re-hosted URLs either way. A value IN the enum is therefore accepted and ignored; send `url` or `b64_json` or omit the field, and the answer is the same. A value OUTSIDE the enum — excluding the UNSET spellings each surface recognises, described next — is REJECTED rather than coerced on BOTH surfaces, and only the shape of the refusal differs: this `/proxy/` route answers a `400` at ingress, and the `POST /v2/models/xai/{model}` Router dispatch answers the per-field `422` it owes the enum published here, before any partner call is made. The two surfaces differ on what counts as UNSET. On this `/proxy/` route an omitted field, an explicit `null` and `\"\"` all mean unset and are coerced like any accepted value. The Router dispatch enforces the enum exactly as published, so `null` and `\"\"` are non-members there like any other: OMIT the field to leave it unset on that surface.","enum":["url","b64_json"],"type":"string"},"size":{"description":"Size of the image (not supported)","type":"string"},"style":{"description":"Style of the image (not supported)","type":"string"},"user":{"description":"A unique identifier representing your end-user, which can help xAI to monitor and detect abuse","type":"string"}},"required":["prompt"],"type":"object"}}}},"responses":{"200":{"description":"OK - the model's native JSON output, returned unchanged; the schema describes the provider's terminal result document as Router returns it.","content":{"application/json":{"schema":{"allOf":[{"$ref":"#/components/schemas/XAIImageGenerationResponse"}],"description":"Comfy Router output schema for the xAI Grok Imagine image models: the `POST /v1/images/generations` document xAI returns, forwarded unchanged EXCEPT for the re-hosted `data[]` entries described below - every other field is xAI's own. This family is DIRECT-RETURN — `routerresult/classification.go` records `{provider: xai, endpoint: /v1/images/generations}` as `ReturnModeDirect` with no poll route — so the body a caller receives is the finished generation on the original call rather than a task handle.\nThe generated images are in `data`, and on THIS surface an entry is addressed by its `url`. Router coerces the outbound `response_format` to `url` for this family and re-hosts every image it can fetch onto Comfy storage, replacing xAI's URL with a signed Comfy one and clearing `b64_json`, so a request that asked for `b64_json` is answered exactly as a `url` request is. `mime_type` names the encoding. The `/proxy/` image routes coerce `response_format` to `url` the same way, so they do not preserve a `b64_json` request either.\nDurability is per entry, not per response: an image whose re-host fails keeps xAI's own answer — under the coerced `response_format` that is xAI's short-lived URL — rather than a Comfy one, and the rest of the response is unaffected. So one `data` array can mix durable Comfy URLs with expiring partner ones, and a caller that stores or replays this document should not assume every URL in it outlives the call. Such a response IS the answer to the call that produced it, returned and charged exactly as a fully re-hosted one is — but it is replayable only BRIEFLY. A prompt retry carrying the same `Idempotency-Key` — within a few minutes of the original, which is where a connection dropped mid-call puts an SDK's automatic re-send — is answered from the record exactly as any other replay is, because the partner's own link is still alive that soon. A retry after that is answered `409 invalid_input` rather than handed a document whose links may already have expired. The key is consumed either way, so no retry ever re-runs or re-charges; once the replay window has passed, use a new key to run the generation again. A response whose every entry re-hosted cleanly replays normally for the full 24 hours, because a Comfy signed URL is minted with the same lifetime the record is kept for.\n`block_reason` and `usage` are populated WITHOUT any `data` when xAI's input moderation refuses the request, so a caller must key completion off the generated asset rather than off a `200` alone. Key it off an ENTRY of `data` carrying a `url` or a `b64_json`, not off `data` itself: a present-but-asset-less entry is still a billed call that produced nothing, and the container test cannot tell the two apart. That is what the nightly SDK case for this family asserts — the leaves `data[].url` and `data[].b64_json` rather than the `data` container (`testing/e2e/router_sdk/cases.d/xai_image.json`).","example":{"data":[{"mime_type":"image/jpeg","url":"https://example.invalid/xai/grok-imagine-image/generated.jpg"},{"mime_type":"image/jpeg","url":"https://example.invalid/xai/grok-imagine-image/generated-2.jpg"}],"usage":{"cost_in_usd_ticks":200000000}}}}}}}}}},"components":{"schemas":{"XAIGeneratedImage":{"description":"A generated image from xAI","properties":{"b64_json":{"description":"A base64-encoded string representation of the generated image in jpeg encoding (if response_format is b64_json)","type":"string"},"mime_type":{"description":"The MIME type of the generated image (e.g. image/png, image/jpeg, image/webp).","type":"string"},"url":{"description":"A url to the generated image (if response_format is url)","type":"string"}},"type":"object"},"XAIImageGenerationResponse":{"description":"Response from xAI image generation or editing","properties":{"block_reason":{"description":"If the request was blocked by input moderation, contains the block reason","type":"string"},"data":{"description":"A list of generated image objects","items":{"$ref":"#/components/schemas/XAIGeneratedImage"},"type":"array"},"usage":{"$ref":"#/components/schemas/XAIImageUsage"}},"type":"object"},"XAIImageUsage":{"description":"Usage information for the image generation request","properties":{"cost_in_usd_ticks":{"description":"Accurate cost of this request in USD ticks (10,000,000,000 ticks = 1 USD)","type":"integer"}},"type":"object"}},"securitySchemes":{"ApiKeyAuth":{"type":"apiKey","in":"header","name":"X-API-Key"},"BearerAuth":{"type":"http","scheme":"bearer","bearerFormat":"JWT"}}},"x-comfy-router-model-id":"xai/grok-imagine-image","x-comfy-input-schema-authored":true,"x-comfy-output-schema-authored":true,"x-comfy-router-input-modalities":["text"],"x-comfy-router-output-modalities":["image"]}
+{"openapi":"3.0.2","info":{"title":"xai/grok-imagine-image","description":"The request body Comfy Router accepts for the model \"xai/grok-imagine-image\", and the response body it returns. The INPUT schema is the same schema the server validates a call against before it reaches the provider, so what is published and what is enforced cannot differ. `x-comfy-router-input-modalities` records the media kinds the model is documented to accept; the request schema is shared across sibling models and may admit media kinds this model ignores. The OUTPUT schema describes the provider's native result document exactly as Router returns it: Router does not validate, narrow or re-envelope the response, so the output schema is descriptive rather than enforced, and Comfy owns no output shape of its own.","version":"a50bbf97594a"},"paths":{"/v2/models/xai/grok-imagine-image":{"post":{"operationId":"runRouterModel","summary":"Run xai/grok-imagine-image synchronously.","security":[{"BearerAuth":[]},{"ApiKeyAuth":[]}],"requestBody":{"required":true,"content":{"application/json":{"schema":{"description":"Request body for xAI Grok Imagine image generation","example":{"n":1,"prompt":"A single red maple leaf on a plain white background."},"properties":{"aspect_ratio":{"default":"auto","description":"Aspect ratio of the generated image. Defaults to auto for automatically selecting the best ratio for the prompt.","enum":["1:1","3:4","4:3","9:16","16:9","2:3","3:2","9:19.5","19.5:9","9:20","20:9","1:2","2:1","auto"],"type":"string"},"model":{"default":"grok-imagine-image","description":"Model to be used. Supported: grok-imagine-image (default), grok-imagine-image-pro, grok-imagine-image-quality, grok-imagine-image-2.0. Deprecated -beta ids are aliased to their GA model.","nullable":true,"type":"string"},"n":{"default":1,"description":"Number of images to be generated","maximum":10,"minimum":1,"type":"integer"},"prompt":{"description":"Prompt for image generation","type":"string"},"quality":{"description":"Quality of the output image. For grok-imagine-image-2.0 this selects the price tier: low, medium, or auto (xAI's default when omitted; auto serves low for generation and medium for edits). Other models ignore it.","enum":["low","medium","high","auto"],"type":"string"},"resolution":{"default":"1k","description":"Resolution of the generated image. Defaults to 1k.","enum":["1k","2k"],"type":"string"},"response_format":{"default":"url","description":"Response format to return the image in. Can be url or b64_json. Comfy coerces this to `url` on the outbound request — on the Comfy Router dispatch (`POST /v2/models/xai/{model}`) and on this `/proxy/` route too — because image results are served as re-hosted URLs either way. A value IN the enum is therefore accepted and ignored; send `url` or `b64_json` or omit the field, and the answer is the same. A value OUTSIDE the enum — excluding the UNSET spellings each surface recognises, described next — is REJECTED rather than coerced on BOTH surfaces, and only the shape of the refusal differs: this `/proxy/` route answers a `400` at ingress, and the `POST /v2/models/xai/{model}` Router dispatch answers the per-field `422` it owes the enum published here, before any partner call is made. The two surfaces differ on what counts as UNSET. On this `/proxy/` route an omitted field, an explicit `null` and `\"\"` all mean unset and are coerced like any accepted value. The Router dispatch enforces the enum exactly as published, so `null` and `\"\"` are non-members there like any other: OMIT the field to leave it unset on that surface.","enum":["url","b64_json"],"type":"string"},"size":{"description":"Size of the image (not supported)","type":"string"},"style":{"description":"Style of the image (not supported)","type":"string"},"user":{"description":"A unique identifier representing your end-user, which can help xAI to monitor and detect abuse","type":"string"}},"required":["prompt"],"type":"object"}}}},"responses":{"200":{"description":"OK - the model's native JSON output, returned unchanged; the schema describes the provider's terminal result document as Router returns it.","content":{"application/json":{"schema":{"allOf":[{"$ref":"#/components/schemas/XAIImageGenerationResponse"}],"description":"Comfy Router output schema for the xAI Grok Imagine image models: the `POST /v1/images/generations` document xAI returns, forwarded unchanged EXCEPT for the re-hosted `data[]` entries described below - every other field is xAI's own. This family is DIRECT-RETURN — `routerresult/classification.go` records `{provider: xai, endpoint: /v1/images/generations}` as `ReturnModeDirect` with no poll route — so the body a caller receives is the finished generation on the original call rather than a task handle.\nThe generated images are in `data`, and on THIS surface an entry is addressed by its `url`. Router coerces the outbound `response_format` to `url` for this family and re-hosts every image it can fetch onto Comfy storage, replacing xAI's URL with a signed Comfy one and clearing `b64_json`, so a request that asked for `b64_json` is answered exactly as a `url` request is. `mime_type` names the encoding. The `/proxy/` image routes coerce `response_format` to `url` the same way, so they do not preserve a `b64_json` request either.\nDurability is per entry, not per response: an image whose re-host fails keeps xAI's own answer — under the coerced `response_format` that is xAI's short-lived URL — rather than a Comfy one, and the rest of the response is unaffected. So one `data` array can mix durable Comfy URLs with expiring partner ones, and a caller that stores or replays this document should not assume every URL in it outlives the call. Such a response IS the answer to the call that produced it, returned and charged exactly as a fully re-hosted one is — but it is replayable only BRIEFLY. A prompt retry carrying the same `Idempotency-Key` — within a few minutes of the original, which is where a connection dropped mid-call puts an SDK's automatic re-send — is answered from the record exactly as any other replay is, because the partner's own link is still alive that soon. A retry after that is answered `409 invalid_input` rather than handed a document whose links may already have expired. The key is consumed either way, so no retry ever re-runs or re-charges; once the replay window has passed, use a new key to run the generation again. A response whose every entry re-hosted cleanly replays normally for the full 24 hours, because a Comfy signed URL is minted with the same lifetime the record is kept for.\n`block_reason` and `usage` are populated WITHOUT any `data` when xAI's input moderation refuses the request, so a caller must key completion off the generated asset rather than off a `200` alone. Key it off an ENTRY of `data` carrying a `url` or a `b64_json`, not off `data` itself: a present-but-asset-less entry is still a billed call that produced nothing, and the container test cannot tell the two apart. That is what the nightly SDK case for this family asserts — the leaves `data[].url` and `data[].b64_json` rather than the `data` container (`testing/e2e/router_sdk/cases.d/xai_image.json`).","example":{"data":[{"mime_type":"image/jpeg","url":"https://example.invalid/xai/grok-imagine-image/generated.jpg"},{"mime_type":"image/jpeg","url":"https://example.invalid/xai/grok-imagine-image/generated-2.jpg"}],"usage":{"cost_in_usd_ticks":200000000}}}}}}}}}},"components":{"schemas":{"XAIGeneratedImage":{"description":"A generated image from xAI","properties":{"b64_json":{"description":"A base64-encoded string representation of the generated image in jpeg encoding (if response_format is b64_json)","type":"string"},"mime_type":{"description":"The MIME type of the generated image (e.g. image/png, image/jpeg, image/webp).","type":"string"},"url":{"description":"A url to the generated image (if response_format is url)","type":"string"}},"type":"object"},"XAIImageGenerationResponse":{"description":"Response from xAI image generation or editing","properties":{"block_reason":{"description":"If the request was blocked by input moderation, contains the block reason","type":"string"},"data":{"description":"A list of generated image objects","items":{"$ref":"#/components/schemas/XAIGeneratedImage"},"type":"array"},"usage":{"$ref":"#/components/schemas/XAIImageUsage"}},"type":"object"},"XAIImageUsage":{"description":"Usage information for the image generation request","properties":{"cost_in_usd_ticks":{"description":"Accurate cost of this request in USD ticks (10,000,000,000 ticks = 1 USD)","type":"integer"}},"type":"object"}},"securitySchemes":{"ApiKeyAuth":{"type":"apiKey","in":"header","name":"X-API-Key"},"BearerAuth":{"type":"http","scheme":"bearer","bearerFormat":"JWT"}}},"x-comfy-router-model-id":"xai/grok-imagine-image","x-comfy-input-schema-authored":true,"x-comfy-output-schema-authored":true,"x-comfy-router-input-modalities":["text"],"x-comfy-router-output-modalities":["image"]}
diff --git a/tutorials/video/minimax/minimax-h3-native.mdx b/tutorials/video/minimax/minimax-h3-native.mdx
index da9c0202ea..237e132aaa 100644
--- a/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/tutorials/video/minimax/minimax-h3-native.mdx
@@ -253,29 +253,31 @@ Connect a mask to the sampler `denoise_mask`: **0** = keep, **1** = regenerate.
 
 ### Long clips with context windows (#17425)
 
-[`MiniMaxH3ContextWindows`](https://github.com/Comfy-Org/ComfyUI/pull/17425) samples a long clip in overlapping windows, so each step holds one window in memory instead of the whole latent. Wire it between `UNETLoader` and `BasicGuider` (model in, model out), the same place as any other model patch.
+[`MiniMaxH3ContextWindows`](https://github.com/Comfy-Org/ComfyUI/pull/17425) samples a long clip in overlapping windows, so each step runs the model on one window instead of the whole clip. Connect it between the turbo LoRA switch (`ComfySwitchNode`) and `BasicGuider`, model in and model out. `BasicScheduler` keeps the switch output and does not need the patch.
 
 | Control | Meaning | Default |
 | --- | --- | --- |
-| `context_length` | Window size in frames, **17n+5** | **124**, the default clip length |
+| `context_length` | Window size in frames, **17n+5** | **124**, the template's 5 s duration snapped to the grid |
 | `context_overlap` | Frames shared by adjacent windows, **17n+5** | **39** |
-| `fuse_method` | How overlapping windows are combined | `pyramid` |
+| `fuse_method` | How overlapping windows are combined: `pyramid`, `relative`, `flat`, `overlap-linear` | `pyramid` |
 | `freenoise` | Shuffles noise across windows to blend them | On |
 
-Windows start on the model's 17-frame chunks, so video and audio stay in step across a boundary. The node keeps `context_overlap` below the window length. At the default **124** the whole 5-second clip is a single window, so the node only changes sampling once the clip runs past the window: lower `context_length` to fit a longer clip on the same card, raise it to give each step more surrounding context at a higher memory cost.
+Windows start on the model's 17-frame chunks, so video and audio stay in step across a boundary. `context_overlap` is capped at the window length minus **17** frames, and a larger value is clamped without a warning. A 5 s clip fits one window at the default **124**, and the first length that opens a second window is **141** frames. Lower `context_length` to cut the memory each step needs. Raise it for more context per step, at a higher cost.
+
+Memory falls per step, not in total. The full latent, the buffers that fuse the windows and the decode still scale with clip length, so a clip that decodes out of memory stays out of memory.
 
 Two limits come with windowed sampling:
 
-- **The prompt runs in full inside every window.** Shot timings are read per window, not across the clip, so a timed cut fires again in each one. Describe one continuous scene and leave timed cuts out.
+- **Every window receives the full prompt.** Write one continuous scene and leave out timed cuts: a shot list with cut times assumes the whole clip is sampled in one pass.
 - **Windows are fused, not joined.** Content that has to hold exactly across a boundary (on-screen text, set detail, a face) can drift. Keep the camera and the subject steady.
 
-That makes context windows a fit for a scene that runs without a cut: a slow push, a pan, or a second-pass upscale of a finished clip. For a clip that cuts between shots, render the shots separately and join them.
+That makes context windows a fit for a scene that runs without a cut: a slow push, a pan, or an unbroken take that runs past the window. For a clip that cuts between shots, render the shots separately and join them.
 
 ### Generating the soundtrack first
 
 Audio and video are denoised together, so a short schedule or a turbo LoRA softens the track along with the picture. Two passes separate them: settle the audio on a small canvas, then condition the video on it.
 
-1. **Cheap audio pass**: Run a text-to-video pass at **32×32** with `Boolean (Enable Lightning LoRA)` off, **40–50** steps, and the duration the final shot needs. Audio latent frames follow clip length, not canvas size, so this pass produces a full-length track for far less than a full-resolution render. Unlink the Resolution Selector and type the size into its `width` and `height` targets; the selector bottoms out at **0.1** Megapixels. Steps live on the `PrimitiveInt` set to **20** inside the subgraph (the one set to **6** is the turbo step count). Only the audio is kept.
+1. **Cheap audio pass**: Run a text-to-video pass at **32×32** with `Boolean (Enable Lightning LoRA)` off, **40–50** steps, and the duration the final shot needs. Audio latent frames follow clip length, not canvas size, so this pass produces a full-length track for far less than a full-resolution render. Unlink the Resolution Selector and type the size into its `width` and `height` targets; the selector bottoms out at **0.1** Megapixels. Steps live on the `PrimitiveInt` set to **20** inside the subgraph. The other `PrimitiveInt` is the turbo step count, which the subgraph widget sets to **8**. Only the audio is kept.
 2. **Save the track**: The subgraph decodes audio inside and exposes only the video, so add a `SaveAudioAdvanced` node inside the subgraph and connect `VAEDecodeAudio` to it.
 3. **Condition the second pass**: In the R2V workflow, add `LoadAudio`, wire it to `ref_audio`, and keep `audio_vae` connected. Cite the track as `<Audio 1>` and state the relationship: `reference` keeps the timbre, rhythm, and delivery without copying the signal, `fully_copy` makes it the complete final track, `partially_copy` keeps part of the timeline or selected layers, and `weak_reference` keeps only broad similarity. The second pass starts from a settled voice and timing instead of inventing speech from nothing. For the identical file, wire `LoadAudio` into `CreateVideo` instead of relying on the prompt.
 
diff --git a/zh/tutorials/video/minimax/minimax-h3-native.mdx b/zh/tutorials/video/minimax/minimax-h3-native.mdx
index 1a01b633ba..87c0efc93d 100644
--- a/zh/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/zh/tutorials/video/minimax/minimax-h3-native.mdx
@@ -2,14 +2,14 @@
 title: "ComfyUI MiniMax H3 文生视频、图生视频与参考生视频工作流"
 description: "ComfyUI 中 MiniMax H3 原生工作流的逐步示例，含帧锚定与潜在噪声蒙版技巧，帮助你对生成视频做更精细的控制。"
 sidebarTitle: "原生工作流"
-translationSourceHash: 2a225f15
+translationSourceHash: b16dd815
 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx
 translationBlockHashes:
   "_intro": dde651e5
   "MiniMax H3 Text to Video (T2V)": 42ec8cfb
   "MiniMax H3 Image to Video (I2V)": 2a06e3fa
   "MiniMax H3 Reference to Video (R2V)": 961829df
-  "Advanced workflows with the native nodes": bbaed645
+  "Advanced workflows with the native nodes": d423a0db
 ---
 本页涵盖三个基础 MiniMax H3 模板及原生节点高级技巧。能力与授权见 [MiniMax H3 概览](/zh/tutorials/video/minimax/minimax-h3)。
 
@@ -260,29 +260,31 @@ ComfyUI/
 
 ### 用上下文窗口生成长片段 (#17425)
 
-[`MiniMaxH3ContextWindows`](https://github.com/Comfy-Org/ComfyUI/pull/17425) 把长片段切成互相重叠的窗口采样，每一步只把一个窗口放进显存，不必放下整个 latent。接在 `UNETLoader` 与 `BasicGuider` 之间（模型进、模型出），与其他模型 patch 的位置相同。
+[`MiniMaxH3ContextWindows`](https://github.com/Comfy-Org/ComfyUI/pull/17425) 把长片段分成互相重叠的窗口来采样，每一步只对一个窗口跑模型，而不是整段片段。接在 turbo LoRA 开关（`ComfySwitchNode`）与 `BasicGuider` 之间，模型进、模型出。`BasicScheduler` 继续用开关的输出，不需要打这个补丁。
 
 | 控件 | 含义 | 默认值 |
 | --- | --- | --- |
-| `context_length` | 窗口长度（帧），**17n+5** | **124**，即默认片段长度 |
+| `context_length` | 窗口长度（帧），**17n+5** | **124**，即模板 5 秒时长对齐到该网格后的帧数 |
 | `context_overlap` | 相邻窗口共享的帧数，**17n+5** | **39** |
-| `fuse_method` | 重叠窗口的融合方式 | `pyramid` |
+| `fuse_method` | 重叠窗口的融合方式：`pyramid`、`relative`、`flat`、`overlap-linear` | `pyramid` |
 | `freenoise` | 跨窗口打乱噪波以衔接 | 开启 |
 
-窗口从模型的 17 帧块边界开始，视频与音频跨边界仍对齐。节点会把 `context_overlap` 压在窗口长度以下。默认 **124** 时整段 5 秒片段就是一个窗口，片段超过窗口长度后节点才真正起作用：调小 `context_length` 让更长的片段塞进同一张卡，调大则每一步拿到更多上下文，显存占用随之上升。
+窗口从模型的 17 帧块边界开始，视频与音频跨边界仍对齐。`context_overlap` 上限是窗口长度减 **17** 帧，设得更大只会被静默压回上限。默认 **124** 下 5 秒片段占一个窗口，第一个能开出第二个窗口的长度是 **141** 帧。调小 `context_length` 降低每一步的显存占用。调大则每一步拿到更多上下文，代价更高。
+
+省下的是每一步的显存，不是总量。整段 latent、融合窗口用的缓冲区和解码仍随片段长度增长，所以解码就会爆显存的片段，改完还是爆。
 
 窗口化采样自带两个限制：
 
-- **每个窗口都会跑完整提示词。** 分镜时间按窗口读，不是按整段读，定时切镜会在每个窗口里再触发一次。写一个连续的场景，不要写定时切镜。
+- **每个窗口都收到完整提示词。** 写一个连续的场景，不要写定时切镜：带切镜时间的分镜假定整段片段只采样一次。
 - **窗口是融合，不是拼接。** 需要跨边界保持一致的内容（画面文字、场景细节、人脸）可能漂移。镜头和主体保持稳定。
 
-所以上下文窗口适合没有切镜的连续场景：缓慢推镜、平移，或对成片做二次放大。需要切镜的片段，分镜单独生成再拼接。
+所以上下文窗口适合没有切镜的连续场景：缓慢推镜、平移，或超过窗口长度的一镜到底。需要切镜的片段，分镜单独生成再拼接。
 
 ### 先生成音轨
 
 音频与视频在同一次去噪中生成，短调度或 turbo LoRA 会连同画面一起把音轨变软。分两遍就能拆开：先在小画布上把音频定下来，再让视频以它为条件。
 
-1. **便宜的音频遍**：以 **32×32** 画布、关闭 `Boolean (Enable Lightning LoRA)`、**40–50** 步、按最终镜头需要的时长跑一遍文生视频。音频 latent 的帧数跟着片段长度走，与画布尺寸无关，所以这一遍能用远低于一次全分辨率渲染的代价产出完整长度的音轨。断开 Resolution Selector，把尺寸直接填进它对应的 `width` 和 `height`；选择器最小只到 **0.1** 百万像素。步数在子图里设为 **20** 的那个 `PrimitiveInt` 上（设为 **6** 的那个是 turbo 步数）。这一遍只保留音频。
+1. **便宜的音频遍**：以 **32×32** 画布、关闭 `Boolean (Enable Lightning LoRA)`、**40–50** 步、按最终镜头需要的时长跑一遍文生视频。音频 latent 的帧数跟着片段长度走，与画布尺寸无关，所以这一遍能用远低于一次全分辨率渲染的代价产出完整长度的音轨。断开 Resolution Selector，把尺寸直接填进它对应的 `width` 和 `height`；选择器最小只到 **0.1** 百万像素。步数在子图里设为 **20** 的那个 `PrimitiveInt` 上。另一个 `PrimitiveInt` 是 turbo 步数，子图控件把它设为 **8**。这一遍只保留音频。
 2. **存下音轨**：子图在内部解码音频，只把视频暴露出来，所以在子图内加一个 `SaveAudioAdvanced`，把 `VAEDecodeAudio` 接到它上面。
 3. **给第二遍做条件**：在 R2V 工作流中加入 `LoadAudio`，接到 `ref_audio`，并保持 `audio_vae` 连接。提示词里用 `<Audio 1>` 引用它，并写明关系：`reference` 只取音色、节奏与语气，不复制信号本身；`fully_copy` 把它当作完整的最终音轨；`partially_copy` 只复制一部分时间线或部分音轨层；`weak_reference` 只保留大类与氛围上的相似。第二遍于是从一个已经确定的音色与节奏起步，而不是从零编造语音。要用完全相同的文件，把 `LoadAudio` 接到 `CreateVideo`，不要依赖提示词。
 

From f9575bdefcd380be28721b36484e866547362448 Mon Sep 17 00:00:00 2001
From: lin-bot23 <linzhixiong23@gmail.com>
Date: Mon, 2 Nov 2026 18:07:31 +0800
Subject: [PATCH 11/13] docs(minimax): cover voice identity, accent and cut-off
 lines on the H3 prompt guide

---
 .../video/minimax/minimax-h3-prompt-guide.mdx         |  9 ++++++---
 .../video/minimax/minimax-h3-prompt-guide.mdx         | 11 +++++++----
 .../video/minimax/minimax-h3-prompt-guide.mdx         | 11 +++++++----
 tutorials/video/minimax/minimax-h3-prompt-guide.mdx   |  5 ++++-
 .../video/minimax/minimax-h3-prompt-guide.mdx         | 11 +++++++----
 5 files changed, 31 insertions(+), 16 deletions(-)

diff --git a/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
index ca36c6a60b..374833b0c0 100644
--- a/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
+++ b/es/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
@@ -2,7 +2,7 @@
 title: "Guía de prompts de ComfyUI MiniMax H3: guías oficiales, consejos y embeddings de prompt"
 description: "Escribe mejores prompts de MiniMax H3 en ComfyUI: guías oficiales, consejos generales, reglas de diálogo y hablantes, y embeddings de estilo."
 sidebarTitle: "Guía de prompts"
-translationSourceHash: cdaa2314
+translationSourceHash: c65880fa
 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx
 translationBlockHashes:
   "_intro": da22cc9d
@@ -10,7 +10,7 @@ translationBlockHashes:
   "General tips": 447ee59b
   "Canvas size": 8c15c46c
   "Shots and cuts": 8fc80fd4
-  "Dialogue and speakers": e403c931
+  "Dialogue and speakers": a91ecba8
   "Prompt embeddings": 9c899169
 ---
 H3 espera planos, cámara, diálogo y SFX en un solo prompt multimodal. Esta página reúne guías oficiales, consejos entre flujos de trabajo, reglas de planos y diálogo, y embeddings de estilo.
@@ -101,8 +101,11 @@ Escribe el movimiento dentro de la acción. No acumules las palabras como etique
 
 ## Diálogo y hablantes
 
-- **IDs de hablante**: `(S1)`, `(S2)`, … para cada voz (incluidas las que están fuera de pantalla y las que cantan). Los IDs indexan hablantes, no números de personaje. Dúo: `(S1,S2)`
+- **IDs de hablante**: `(S1)`, `(S2)`, … para cada voz (incluidas las que están fuera de pantalla y las que cantan). Los IDs indexan hablantes, no números de personaje, y un hablante conserva el mismo ID en todos los planos. Varios hablan o cantan a la vez: `(S1,S2)`. Los personajes que nunca vocalizan no llevan ID
 - **Contenido de `<d>`**: el ID, la acción y la forma de decirlo van fuera de las etiquetas; dentro de las etiquetas, la etiqueta de idioma + las palabras pronunciadas textualmente, en el párrafo del plano donde se dicen
+- **Identidad de la voz**: en la primera aparición, da fuera de la etiqueta la información suficiente para fijar la voz: tipo de personaje, edad, género, si está en pantalla, tono, timbre, velocidad al hablar y acento. Usa los atributos que pida el plano
+- **Acento y referencias de audio**: el acento va en la descripción, no en la etiqueta `[English]` (`(S2) answers in a low cockney voice`). Con una referencia de audio, nómbrala en la misma frase: `answering in the voice referenced from <Audio 1>`
+- **Diálogo truncado**: usa `<cutoff>` en el punto donde el final del clip corta la línea
 - **Puntuación en `<d>`**: conserva `, . ? !`; elimina repeticiones decorativas y emoji; usa `[unclear]` cuando no estés seguro. Los puntos suspensivos suelen perjudicar la inteligibilidad. Cierra cada línea con un solo signo final antes de `</d>`, sin signos repetidos (`?!?`) y sin nada después de la etiqueta (`ready?</d>.`)
 - **Análisis de retención (R2V)**: describe qué conserva cada referencia, no lo que se dice. No escribas diálogo ni letras ahí, ni IDs `(Sx)`. Cada línea hablada va solo dentro de `<d>` en el plano donde se oye
 - **Referencias solo de timbre**: cuando `<Audio N>` aporta solo timbre, ritmo, emoción o forma de decirlo, no reutilices sus palabras. Escribe la línea nueva en `<d>` y marca el audio como `reference`, por ejemplo `<Audio 2>: reference - the target speaker follows <Audio 2>'s voice timbre and measured delivery without copying the original signal`
diff --git a/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
index 5f32b8f5a6..3903a99455 100644
--- a/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
+++ b/ja/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
@@ -2,15 +2,15 @@
 title: "ComfyUI MiniMax H3 プロンプトガイド: 公式ガイド、ヒント、プロンプト埋め込み"
 description: "ComfyUI でより良い MiniMax H3 プロンプトを書く方法: MiniMax の公式プロンプト作成ガイド、一般的なプロンプトのコツ、セリフと話者のルール、スタイル埋め込み。"
 sidebarTitle: "プロンプトガイド"
-translationSourceHash: cdaa2314
+translationSourceHash: c65880fa
 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx
 translationBlockHashes:
   "_intro": da22cc9d
   "Official guides": 3d762c67
   "General tips": 447ee59b
   "Canvas size": 8c15c46c
-  "Shots and cuts": 8184dae6
-  "Dialogue and speakers": 4985438e
+  "Shots and cuts": 8fc80fd4
+  "Dialogue and speakers": a91ecba8
   "Prompt embeddings": 9c899169
 ---
 
@@ -103,8 +103,11 @@ non_diegetic_music: ...
 
 ## セリフと話者
 
-- **話者 ID**: `(S1)`、`(S2)`…（オフ画面/歌唱含む）; デュエット `(S1,S2)`
+- **話者 ID**: `(S1)`、`(S2)`…（オフ画面/歌唱含む）; ID は話者単位で振り、同じ話者はショットをまたいでも同じ ID を使う; 同時に話す・歌う `(S1,S2)`; 発声しないキャラには ID を付けない
 - **`<d>`**: ID・動作・演技はタグ外; タグ内は言語タグと原文
+- **声の同一性**: 初登場時に、声を固定する情報をタグ外に書く（人物の種類・年齢・性別・画面内か画面外か・音高・音色・話速・アクセントのうち、ショットに必要な項目）
+- **アクセントと音声参照**: アクセントは説明文に書き、`[English]` タグには入れない（`(S2) answers in a low cockney voice`）; 音声参照がある場合は同じ文で名指す（`answering in the voice referenced from <Audio 1>`）
+- **途中で切れるセリフ**: 動画の終端でセリフが途切れる場合、途切れた位置に `<cutoff>` を置く
 - **`<d>` 内句読点**: `, . ? !` を保持; 装飾的な繰り返しや絵文字は削除; 不明は `[unclear]`; 省略号は避ける; 各行は `</d>` の前に文末記号を 1 つだけ置き、重ねた記号（`?!?`）や閉じタグの後ろの記号（`ready?</d>.`）は置かない
 - **retention_analysis（R2V）**: 参照が何を保持したかだけを書き、セリフや歌詞、`(Sx)` は置かない; セリフはそれが聞こえるショットの `<d>` 内にのみ置く
 - **音色のみの参照**: `<Audio N>` が音色・リズム・感情・演技だけを提供する場合はその言葉を使わない; 新しいセリフを `<d>` に書き、音声は `reference` とする（例: `<Audio 2>: reference - the target speaker follows <Audio 2>'s voice timbre and measured delivery without copying the original signal`）
diff --git a/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
index 15423caf8d..8166db07eb 100644
--- a/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
+++ b/ko/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
@@ -2,15 +2,15 @@
 title: "ComfyUI MiniMax H3 프롬프트 가이드: 공식 가이드, 팁, 프롬프트 임베딩"
 description: "ComfyUI에서 더 나은 MiniMax H3 프롬프트 작성하기: MiniMax 공식 프롬프트 작성 가이드, 일반 프롬프트 팁, 대사와 화자 규칙, 스타일 임베딩."
 sidebarTitle: "프롬프트 가이드"
-translationSourceHash: cdaa2314
+translationSourceHash: c65880fa
 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx
 translationBlockHashes:
   "_intro": da22cc9d
   "Official guides": 3d762c67
   "General tips": 447ee59b
   "Canvas size": 8c15c46c
-  "Shots and cuts": 8184dae6
-  "Dialogue and speakers": 4985438e
+  "Shots and cuts": 8fc80fd4
+  "Dialogue and speakers": a91ecba8
   "Prompt embeddings": 9c899169
 ---
 
@@ -103,8 +103,11 @@ non_diegetic_music: ...
 
 ## 대사와 화자
 
-- **화자 ID**: `(S1)`, `(S2)`…(오프스크린/노래 포함); 듀엣 `(S1,S2)`
+- **화자 ID**: `(S1)`, `(S2)`…(오프스크린/노래 포함); ID는 화자 단위로 붙이고 같은 화자는 샷이 바뀌어도 같은 ID를 쓴다; 함께 말하거나 노래할 때 `(S1,S2)`; 소리 내지 않는 캐릭터에는 ID를 주지 않는다
 - **`<d>`**: ID·동작·연기는 태그 밖; 태그 안 언어 태그+원문
+- **목소리 정체성**: 처음 등장할 때 목소리를 고정할 정보를 태그 밖에 적는다(캐릭터 유형·나이·성별·화면 안팎 여부·음높이·음색·말 속도·억양 중 샷에 필요한 항목)
+- **억양과 오디오 참조**: 억양은 설명에 적고 `[English]` 태그에 넣지 않는다 (`(S2) answers in a low cockney voice`); 오디오 참조가 있으면 같은 문장에서 지칭한다 (`answering in the voice referenced from <Audio 1>`)
+- **중간에 끊기는 대사**: 영상 끝에서 대사가 잘리면 잘린 지점에 `<cutoff>`를 둔다
 - **`<d>` 문장부호**: `, . ? !` 유지; 장식용 반복과 이모지 제거; 불명 `[unclear]`; 말줄임 자제; 각 줄은 `</d>` 앞에 문장 끝 부호를 하나만 두고, 겹친 부호(`?!?`)나 닫는 태그 뒤의 부호(`ready?</d>.`)는 쓰지 않는다
 - **retention_analysis(R2V)**: 참조가 무엇을 유지했는지만 적고 대사·가사·`(Sx)`는 넣지 않는다; 대사는 그것이 들리는 샷의 `<d>` 안에만 둔다
 - **음색만 참조**: `<Audio N>`이 음색·리듬·감정·전달만 제공할 때는 그 말을 그대로 쓰지 않는다; 새 대사를 `<d>`에 쓰고 오디오는 `reference`로 표시한다 (예: `<Audio 2>: reference - the target speaker follows <Audio 2>'s voice timbre and measured delivery without copying the original signal`)
diff --git a/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
index 0c56a71716..9fbfe98591 100644
--- a/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
+++ b/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
@@ -92,8 +92,11 @@ Write the move into the action. Do not stack the words as labels at the end of t
 
 ## Dialogue and speakers
 
-- **Speaker IDs**: `(S1)`, `(S2)`, … for every voice (including off-screen/singing). IDs index speakers, not character numbers. Duet: `(S1,S2)`
+- **Speaker IDs**: `(S1)`, `(S2)`, … for every voice (including off-screen/singing). IDs index speakers, not character numbers, and a speaker keeps the same ID across shots. Speaking or singing together: `(S1,S2)`. Characters who never vocalize get no ID
 - **`<d>` content**: ID, action, and delivery outside tags; inside tags, language tag + spoken words verbatim, in the shot paragraph where spoken
+- **Voice identity**: On first appearance, give enough outside the tag to fix the voice: character type, age, gender, on-screen or off-screen, pitch, timbre, speaking rate, accent. Use the attributes the shot needs
+- **Accent and audio references**: Write the accent into the identifying phrase, not the `[English]` tag: `(S2) answers in a low cockney voice`. With a voice reference, name it in the same phrase: `answering in the voice referenced from <Audio 1>`
+- **Cut-off lines**: Use `<cutoff>` at the point where the clip ends mid-line
 - **Punctuation in `<d>`**: Keep `, . ? !`; drop decorative repeats and emoji; use `[unclear]` when unsure. Ellipses often hurt intelligibility. Close each line with one terminal mark before `</d>`, with no repeated marks (`?!?`) and nothing after the tag (`ready?</d>.`)
 - **Retention analysis (R2V)**: It records what each reference keeps, not what is said. Write no dialogue or lyrics there and no `(Sx)` IDs. Every spoken line stays inside `<d>` in the shot where it is heard
 - **Timbre-only audio references**: When `<Audio N>` supplies only timbre, rhythm, emotion, or delivery, do not reuse its words. Write the new line in `<d>` and mark the audio `reference`, for example `<Audio 2>: reference - the target speaker follows <Audio 2>'s voice timbre and measured delivery without copying the original signal`
diff --git a/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx b/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
index 81631c3c84..6cdab3e1e2 100644
--- a/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
+++ b/zh/tutorials/video/minimax/minimax-h3-prompt-guide.mdx
@@ -2,15 +2,15 @@
 title: "ComfyUI MiniMax H3 提示词指南：官方指南、技巧与提示词嵌入"
 description: "在 ComfyUI 中写出更好的 MiniMax H3 提示词：MiniMax 官方提示词编写指南、通用提示词技巧、对白与说话人规则，以及风格嵌入。"
 sidebarTitle: "提示词指南"
-translationSourceHash: cdaa2314
+translationSourceHash: c65880fa
 translationFrom: tutorials/video/minimax/minimax-h3-prompt-guide.mdx
 translationBlockHashes:
   "_intro": da22cc9d
   "Official guides": 3d762c67
   "General tips": 447ee59b
   "Canvas size": 8c15c46c
-  "Shots and cuts": 8184dae6
-  "Dialogue and speakers": 4985438e
+  "Shots and cuts": 8fc80fd4
+  "Dialogue and speakers": a91ecba8
   "Prompt embeddings": 9c899169
 ---
 
@@ -103,8 +103,11 @@ non_diegetic_music: ...
 
 ## 对白与说话人
 
-- **说话人 ID**：`(S1)`、`(S2)`…（含画外/演唱）；合唱 `(S1,S2)`
+- **说话人 ID**：`(S1)`、`(S2)`…（含画外/演唱）；ID 按说话人编号，同一说话人跨镜头沿用；同时说或唱 `(S1,S2)`；不出声的角色不给 ID
 - **`<d>`**：ID、动作、语气在标签外；标签内为语言标记与原文
+- **声音特征**：角色首次出场时，在标签外给出足以固定声音的信息：角色类型、年龄、性别、画内或画外、音高、音色、语速、口音，按镜头需要选用
+- **口音与音频参考**：口音写进描述，不写进 `[English]` 标签（`(S2) answers in a low cockney voice`）；有音频参考时在同一句点名（`answering in the voice referenced from <Audio 1>`）
+- **中途截断**：台词被片尾截断时，在截断处写 `<cutoff>`
 - **标点**：保留 `, . ? !`；去掉装饰性重复与 emoji；听不清写 `[unclear]`；慎用省略号；每句在 `</d>` 前以一个句末标点收尾，不连写（`?!?`），也不在结束标签后加标点（`ready?</d>.`）
 - **保留分析（R2V）**：只写参考保留了什么，不写台词；`retention_analysis` 中不放对白、歌词和 `(Sx)`；每句台词只放在听到它的那个镜头的 `<d>` 内
 - **仅音色参考**：`<Audio N>` 只提供音色、节奏、情绪或语气时不沿用其原话；新台词写进 `<d>`，音频标为 `reference`，例如 `<Audio 2>: reference - the target speaker follows <Audio 2>'s voice timbre and measured delivery without copying the original signal`

From 1dfbe9a79ce92ba5d2c54df4dfc891234dd64d97 Mon Sep 17 00:00:00 2001
From: lin-bot23 <linzhixiong23@gmail.com>
Date: Tue, 3 Nov 2026 18:41:37 +0800
Subject: [PATCH 12/13] docs(minimax): cover the H3 sparse attention node and
 the two-pass window tradeoff

---
 .../video/minimax/minimax-h3-native.mdx       | 26 ++++++++++++++++---
 .../video/minimax/minimax-h3-native.mdx       | 26 ++++++++++++++++---
 .../video/minimax/minimax-h3-native.mdx       | 26 ++++++++++++++++---
 tutorials/video/minimax/minimax-h3-native.mdx | 22 +++++++++++++++-
 .../video/minimax/minimax-h3-native.mdx       | 26 ++++++++++++++++---
 5 files changed, 113 insertions(+), 13 deletions(-)

diff --git a/es/tutorials/video/minimax/minimax-h3-native.mdx b/es/tutorials/video/minimax/minimax-h3-native.mdx
index bc76d51183..f83718c07a 100644
--- a/es/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/es/tutorials/video/minimax/minimax-h3-native.mdx
@@ -2,14 +2,14 @@
 title: "Flujos de trabajo de MiniMax H3 en ComfyUI: texto a video, imagen a video y referencia a video"
 description: "Ejemplos paso a paso de flujos nativos de MiniMax H3 en ComfyUI, con anclaje de guías y máscaras de ruido latente para un control más fino del video."
 sidebarTitle: "Flujos nativos"
-translationSourceHash: b16dd815
+translationSourceHash: 48763c0c
 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx
 translationBlockHashes:
   "_intro": dde651e5
   "MiniMax H3 Text to Video (T2V)": 42ec8cfb
   "MiniMax H3 Image to Video (I2V)": 2a06e3fa
   "MiniMax H3 Reference to Video (R2V)": 961829df
-  "Advanced workflows with the native nodes": d423a0db
+  "Advanced workflows with the native nodes": 2ab2df9d
 ---
 Esta página cubre las tres plantillas base de MiniMax H3 y las técnicas avanzadas con nodos nativos. Para conocer las capacidades y el licenciamiento, consulta la [descripción general de MiniMax H3](/es/tutorials/video/minimax/minimax-h3).
 
@@ -271,11 +271,13 @@ Conecta una máscara a `denoise_mask` del sampler: **0** = mantener, **1** = reg
 
 Las ventanas empiezan en los bloques de 17 frames del modelo, de modo que video y audio siguen alineados al cruzar un límite. `context_overlap` tiene como tope la longitud de la ventana menos **17** frames, y un valor mayor se recorta sin aviso. Con el valor por defecto **124** un clip de 5 s ocupa una sola ventana, y la primera longitud que abre una segunda ventana es **141** frames. Baja `context_length` para reducir la memoria de cada paso. Súbelo para dar más contexto a cada paso, con un coste mayor.
 
+Un clip largo sale más barato en dos pasadas. Renderiza una pasada a baja resolución con ventanas largas para asentar movimiento y sonido, y después otra a mayor resolución con ventanas cortas para añadir detalle. La longitud de ventana se cambia por fidelidad al prompt: cada ventana se muestrea como un solo plano, y cuanto más larga es, antes deja el clip de seguir el prompt.
+
 La memoria baja por paso, no en total. El latente completo, los búferes que fusionan las ventanas y el decodificado siguen creciendo con la longitud del clip, así que un clip que se queda sin memoria al decodificar sigue igual.
 
 El muestreo por ventanas trae dos límites:
 
-- **Cada ventana recibe el prompt completo.** Describe una escena continua y deja fuera los cortes con tiempo: un desglose de planos con marcas de tiempo da por hecho que el clip se muestrea de una vez.
+- **Cada ventana recibe el prompt completo.** Con ventanas activas, describe una escena continua en lugar de un desglose de planos con marcas de tiempo: todas las ventanas ven el prompt entero, así que los tiempos de corte ya no encajan.
 - **Las ventanas se fusionan, no se empalman.** El contenido que debe mantenerse idéntico al cruzar un límite (texto en pantalla, detalle del decorado, una cara) puede derivar. Mantén estables la cámara y el sujeto.
 
 Por eso las ventanas de contexto encajan en una escena sin cortes: un push lento, un paneo o un plano sin corte que supera la ventana. Para un clip que cambia de plano, genera los planos por separado y únelos.
@@ -289,3 +291,21 @@ El audio y el video se desruidizan juntos, así que una programación corta o un
 3. **Condicionar la segunda pasada**: En el flujo R2V añade `LoadAudio`, conéctalo a `ref_audio` y mantén `audio_vae` conectado. Cita la pista como `<Audio 1>` e indica la relación: `reference` conserva timbre, ritmo y dicción sin copiar la señal, `fully_copy` la convierte en la pista final completa, `partially_copy` conserva parte de la línea de tiempo o capas seleccionadas, y `weak_reference` solo mantiene una similitud amplia. La segunda pasada parte así de una voz y un ritmo ya resueltos en lugar de inventar el habla desde cero. Para usar el mismo archivo, conecta `LoadAudio` a `CreateVideo` en vez de confiar en el prompt.
 
 El tamaño del lienzo intercambia calidad de voz por tiempo de render. Una pasada a 32×32 es casi gratis, pero a ese tamaño la voz puede sonar metálica. Sube el lienzo cuando la voz tenga que sostenerse en el corte final, y cuenta con que la pasada cueste más. El habla y las letras suelen sobrevivir al lienzo pequeño. Los efectos de sonido y la música atados a la acción en pantalla no, así que genéralos en la segunda pasada y combínalos con `partially_copy`.
+
+### Muestreo más rápido con Model Sparse Attention (#16072)
+
+[`Model Sparse Attention`](https://github.com/Comfy-Org/ComfyUI/pull/16072) sustituye la atención completa por atención de bloques dispersos. Conéctalo en la línea del modelo, entre el interruptor del LoRA turbo (`ComfySwitchNode`) y `BasicGuider`, con el modelo como entrada y como salida. Necesita ComfyUI **0.35.0+**, CUDA y una GPU para la que esté compilado el kernel `sol-attn`. En cualquier otro caso la llamada vuelve a atención densa y el motivo queda en el registro con `verbose`. Las secuencias cortas suelen ir más rápido en denso, así que la ganancia crece con el lienzo y la longitud del clip.
+
+| `method` | Pesos que necesita |
+| --- | --- |
+| `sol-attn` | Ninguno. Funciona en cualquier checkpoint H3 |
+| `sla` | Un modelo destilado con esta selección. Un checkpoint H3 base no sirve |
+| `vsa` | Checkpoints FastH3-VSA del flujo de [aceleración FastVideo](/es/tutorials/video/minimax/minimax-h3-fastvideo) |
+
+`start_percent` y `end_percent` delimitan dónde se aplica la dispersión. Fuera de esa ventana la atención vuelve a ser densa, y el nodo viene con `start_percent` en **0.2** y `end_percent` en **1.0**, lo que deja dispersos los últimos pasos. Pon el final en **0.9** para que el cierre se desruide a atención completa, y adelanta el inicio por debajo de 0.2 cuando quieras más ahorro: la parte media del muestreo aguanta mejor la dispersión que los extremos.
+
+`tau` (por defecto **1.3**) es el deslizador de velocidad y calidad de `sol-attn`. Más alto es más disperso: **1.0** mantiene exactos alrededor del 16% de los bloques clave, **1.5** un 7%, **2.0** un 2.7%. No hay un valor óptimo, solo un punto en ese intercambio. `sla` y `vsa` cambian `tau` por `keep_percent` (por defecto **10**), la proporción de bloques clave que cada bloque de consulta conserva exactos.
+
+`min_tokens` (**12288**) deja densas las secuencias más cortas sin avisar. Con el lienzo 16:9 de la plantilla, **864×480**, un clip de 5 s reúne unos **15k** tokens y sí se dispersa, mientras que uno de 3 s ronda los **9k** y no gana nada hasta que bajes `min_tokens`. La cuenta es por pasada hacia delante, así que con las ventanas de contexto conectadas pasa a ser por ventana, y una pasada de baja resolución puede quedarse por debajo del umbral. Activa `verbose` para registrar si cada forma se ejecutó en disperso y por qué se quedó en denso.
+
+`sink_conditioning` mantiene exactas las filas empaquetadas de texto, audio y referencias de H3. El valor por defecto `exact_kv_and_rows` también ejecuta en denso las filas de consulta del audio generado, así que la banda sonora se sostiene mientras la atención del video va dispersa.
diff --git a/ja/tutorials/video/minimax/minimax-h3-native.mdx b/ja/tutorials/video/minimax/minimax-h3-native.mdx
index c07961e714..9eb00b8b29 100644
--- a/ja/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/ja/tutorials/video/minimax/minimax-h3-native.mdx
@@ -2,14 +2,14 @@
 title: "ComfyUI MiniMax H3 テキストから動画・画像から動画・参照から動画のワークフロー"
 description: "ComfyUI における MiniMax H3 ネイティブワークフローの手順例。ガイドのアンカー固定や latent ノイズマスクで生成動画をより細かく制御できます。"
 sidebarTitle: "ネイティブワークフロー"
-translationSourceHash: b16dd815
+translationSourceHash: 48763c0c
 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx
 translationBlockHashes:
   "_intro": dde651e5
   "MiniMax H3 Text to Video (T2V)": 42ec8cfb
   "MiniMax H3 Image to Video (I2V)": 2a06e3fa
   "MiniMax H3 Reference to Video (R2V)": 961829df
-  "Advanced workflows with the native nodes": d423a0db
+  "Advanced workflows with the native nodes": 2ab2df9d
 ---
 
 このページは 3 つの基本 MiniMax H3 テンプレートとネイティブノードの高度なテクニックを扱います。機能とライセンスは [MiniMax H3 概要](/ja/tutorials/video/minimax/minimax-h3) を参照してください。
@@ -272,11 +272,13 @@ H3 ノードの `positive` と latent を `MiniMaxH3AddGuide` に接続:
 
 ウィンドウはモデルの 17 フレーム塊の境界から始まるため、動画と音声は境界をまたいでも揃います。`context_overlap` の上限はウィンドウ長から **17** フレームを引いた値で、それより大きく設定しても警告なしに切り詰められます。デフォルトの **124** では 5 秒のクリップが 1 ウィンドウに収まり、2 つ目のウィンドウが開く最小の長さは **141** フレームです。`context_length` を下げると 1 ステップあたりのメモリが減ります。上げると 1 ステップあたりの前後文脈が増え、コストも上がります。
 
+長いクリップは 2 パスに分けるほうが安く済みます。低解像度で長いウィンドウを使い、まず動きと音を固めます。そのあと高解像度で短いウィンドウを使い、ディテールを足します。ウィンドウ長とプロンプト追従は交換関係にあり、各ウィンドウは 1 ショットとしてサンプリングされるため、長くするほどクリップがプロンプトから外れるのが早まります。
+
 減るのは 1 ステップあたりのメモリで、総量ではありません。全体の latent、ウィンドウ融合用のバッファ、デコードはクリップ長に比例したままなので、デコードで OOM になるクリップはそのまま OOM になります。
 
 ウィンドウ分割には 2 つの制約があります:
 
-- **どのウィンドウにもプロンプト全体が渡ります。** 1 つの連続したシーンとして書き、時刻指定のカットは入れません。時刻付きのカット割りは、クリップ全体を 1 回でサンプリングする前提の書き方です。
+- **どのウィンドウにもプロンプト全体が渡ります。** ウィンドウを使うときは、時刻付きのカット割りではなく 1 つの連続したシーンとして書いてください。どのウィンドウもプロンプト全体を見るため、カット時刻は噛み合いません。
 - **ウィンドウは融合されるのであって連結ではありません。** 境界をまたいで厳密に保つ必要がある内容（画面内の文字、セットの細部、顔）はドリフトし得ます。カメラと被写体は安定させます。
 
 そのためコンテキストウィンドウは、カットのない連続したシーン（ゆっくりしたプッシュイン、パン、ウィンドウ長を超えるワンカット）に向きます。ショットを切り替えるクリップは、ショットごとに生成してからつなぎます。
@@ -290,3 +292,21 @@ H3 ノードの `positive` と latent を `MiniMaxH3AddGuide` に接続:
 3. **2 パス目を条件づける**: R2V ワークフローに `LoadAudio` を追加して `ref_audio` に接続し、`audio_vae` は接続したままにします。プロンプトでは `<Audio 1>` として参照し、関係を明記します。`reference` は信号をコピーせず音色・リズム・話し方だけを参照し、`fully_copy` は完全な最終トラックとして扱い、`partially_copy` はタイムラインの一部または一部の音声レイヤーだけをコピーし、`weak_reference` はカテゴリや雰囲気の広い類似だけを保ちます。2 パス目はゼロから音声を作るのではなく、確定した声とタイミングから始まります。まったく同じファイルを使う場合は、プロンプトに頼らず `LoadAudio` を `CreateVideo` に接続します。
 
 キャンバスサイズは音色とレンダー時間の交換になります。32×32 のパスはほぼ無料ですが、そのサイズでは声に機械的な響きが乗ることがあります。声を最終カットで保たせたいときはキャンバスを上げてください。そのぶんパスのコストも上がります。台詞や歌詞は小さいキャンバスでも概ね保ちます。画面の動きに合わせた効果音やスコアは保たないので、2 パス目で生成し、`partially_copy` で 1 パス目のトラックと結合します。
+
+### Model Sparse Attention で高速化する（#16072）
+
+[`Model Sparse Attention`](https://github.com/Comfy-Org/ComfyUI/pull/16072) は全アテンションをブロックスパースアテンションに置き換えます。モデル線上の turbo LoRA スイッチ（`ComfySwitchNode`）と `BasicGuider` の間に、モデル入力とモデル出力で接続します。ComfyUI **0.35.0+**、CUDA、そして `sol-attn` カーネルがコンパイル済みの GPU が必要です。満たさない呼び出しは密アテンションに戻り、理由は `verbose` のログに出ます。短いシーケンスは密のほうが速いことが多いため、キャンバスとクリップが長いほど効果が出ます。
+
+| `method` | 必要な重み |
+| --- | --- |
+| `sol-attn` | 不要。どの H3 チェックポイントでも動きます |
+| `sla` | この選択方式で蒸留されたモデル。ベースの H3 チェックポイントでは動きません |
+| `vsa` | [FastVideo 高速化](/ja/tutorials/video/minimax/minimax-h3-fastvideo)ワークフローの FastH3-VSA チェックポイント |
+
+`start_percent` と `end_percent` がスパース化する区間を決めます。区間の外は密アテンションです。`start_percent` の既定値は **0.2**、`end_percent` は **1.0** で、最後のステップまでスパースのままになります。終了点を **0.9** にすると収束前の数ステップが密で回り、さらに節約したいときは開始点を 0.2 より前にします。スケジュールの中盤は両端よりスパース化に耐えます。
+
+`tau`（既定 **1.3**）は `sol-attn` の速度と品質のスライダーです。大きいほどスパースになり、**1.0** でキーブロックの約 16%、**1.5** で約 7%、**2.0** で約 2.7% を正確に保ちます。最適値はなく、このトレードオフ上の一点を選ぶだけです。`sla` と `vsa` では `tau` の代わりに `keep_percent`（既定 **10**）を使い、各クエリブロックが正確に保つキーブロックの割合を指定します。
+
+`min_tokens`（既定 **12288**）未満のシーケンスは、何も表示せず密のまま実行されます。16:9 テンプレートのキャンバス **864×480** では 5 秒クリップが約 **15k** トークンに達してスパースになり、3 秒クリップは **9k** 前後で `min_tokens` を下げるまで効果が出ません。この数は 1 回のフォワードパスあたりなので、コンテキストウィンドウを接続するとウィンドウ単位になり、解像度の低いパスはしきい値を下回ることがあります。`verbose` を有効にすると、各形状がスパースで走ったか、密のままだった理由がログに出ます。
+
+`sink_conditioning` は H3 がまとめたテキスト・音声・参照の行を正確に保ちます。既定の `exact_kv_and_rows` は生成音声のクエリ行も密で回すため、映像側をスパース化してもサウンドトラックは劣化しません。
diff --git a/ko/tutorials/video/minimax/minimax-h3-native.mdx b/ko/tutorials/video/minimax/minimax-h3-native.mdx
index 6f373141df..bfd8bb8c9d 100644
--- a/ko/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/ko/tutorials/video/minimax/minimax-h3-native.mdx
@@ -2,14 +2,14 @@
 title: "ComfyUI MiniMax H3 텍스트·이미지·레퍼런스 기반 비디오 워크플로"
 description: "ComfyUI의 MiniMax H3 네이티브 워크플로 단계별 예제입니다. 가이드 고정과 잠재 노이즈 마스크로 생성 비디오를 더 세밀하게 제어할 수 있습니다."
 sidebarTitle: "네이티브 워크플로"
-translationSourceHash: b16dd815
+translationSourceHash: 48763c0c
 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx
 translationBlockHashes:
   "_intro": dde651e5
   "MiniMax H3 Text to Video (T2V)": 42ec8cfb
   "MiniMax H3 Image to Video (I2V)": 2a06e3fa
   "MiniMax H3 Reference to Video (R2V)": 961829df
-  "Advanced workflows with the native nodes": d423a0db
+  "Advanced workflows with the native nodes": 2ab2df9d
 ---
 
 이 페이지는 세 가지 기본 MiniMax H3 템플릿과 네이티브 노드 고급 기법을 다룹니다. 기능과 라이선스는 [MiniMax H3 개요](/ko/tutorials/video/minimax/minimax-h3)를 참조하세요.
@@ -272,11 +272,13 @@ H3 노드 `positive`·latent를 `MiniMaxH3AddGuide`에 연결:
 
 윈도우는 모델의 17프레임 덩어리 경계에서 시작하므로 영상과 오디오가 경계를 넘어도 어긋나지 않습니다. `context_overlap` 상한은 윈도우 길이에서 **17**프레임을 뺀 값이고, 더 크게 넣으면 경고 없이 잘립니다. 기본값 **124**에서는 5초 클립이 윈도우 하나에 들어가고, 두 번째 윈도우가 열리는 최소 길이는 **141**프레임입니다. `context_length`를 낮추면 단계마다 쓰는 메모리가 줄어듭니다. 높이면 단계마다 더 많은 앞뒤 맥락을 보는 대신 비용이 커집니다.
 
+긴 클립은 두 패스로 나누는 편이 저렴합니다. 낮은 해상도에서 긴 윈도우로 먼저 움직임과 소리를 잡고, 그다음 높은 해상도에서 짧은 윈도우로 세부를 더합니다. 윈도우 길이는 프롬프트 준수와 맞바꾸는 값이라, 각 윈도우가 한 숏으로 샘플링되므로 길수록 클립이 프롬프트에서 일찍 벗어납니다.
+
 줄어드는 것은 단계당 메모리이지 총량이 아닙니다. 전체 latent, 윈도우를 합치는 버퍼, 디코드는 클립 길이에 비례하므로, 디코드에서 메모리가 부족한 클립은 그대로 부족합니다.
 
 윈도우 분할에는 제약이 두 가지 있습니다:
 
-- **모든 윈도우가 프롬프트 전체를 받습니다.** 하나의 연속된 장면으로 쓰고 시각 지정 컷은 넣지 않습니다. 시각이 붙은 컷 구성은 클립 전체를 한 번에 샘플링한다는 전제로 쓰인 것입니다.
+- **모든 윈도우가 프롬프트 전체를 받습니다.** 윈도우를 켜면 시각이 붙은 컷 구성 대신 하나의 연속된 장면으로 쓰세요. 모든 윈도우가 프롬프트 전체를 보므로 컷 시각이 맞지 않습니다.
 - **윈도우는 이어 붙이는 게 아니라 합쳐집니다.** 경계를 넘어 정확히 유지해야 하는 내용(화면 속 글자, 세트 디테일, 얼굴)은 밀릴 수 있습니다. 카메라와 피사체는 안정적으로 유지합니다.
 
 그래서 컨텍스트 윈도우는 컷이 없는 연속 장면에 맞습니다. 느린 푸시인, 팬, 윈도우보다 긴 원테이크 같은 경우입니다. 샷을 전환하는 클립은 샷별로 생성한 뒤 이어 붙입니다.
@@ -290,3 +292,21 @@ H3 노드 `positive`·latent를 `MiniMaxH3AddGuide`에 연결:
 3. **두 번째 패스 조건 걸기**: R2V 워크플로에 `LoadAudio`를 추가해 `ref_audio`에 연결하고 `audio_vae`는 연결된 상태로 둡니다. 프롬프트에서 `<Audio 1>`로 참조하고 관계를 명시하세요. `reference`는 신호를 복사하지 않고 음색, 리듬, 말투만 참조하고, `fully_copy`는 완전한 최종 트랙으로 취급하며, `partially_copy`는 타임라인의 일부나 일부 오디오 레이어만 복사하고, `weak_reference`는 범주와 분위기의 넓은 유사성만 유지합니다. 두 번째 패스는 빈 상태에서 음성을 지어내는 대신 확정된 목소리와 타이밍에서 시작합니다. 완전히 같은 파일을 쓰려면 프롬프트에 의존하지 말고 `LoadAudio`를 `CreateVideo`에 연결하세요.
 
 캔버스 크기는 음질과 렌더 시간을 맞바꿉니다. 32×32 패스는 거의 무료지만, 그 크기에서는 목소리에 기계적인 느낌이 붙을 수 있습니다. 목소리가 최종 컷에서 버텨야 한다면 캔버스를 올리세요. 그만큼 패스 비용도 올라갑니다. 대사와 가사는 작은 캔버스에서도 대체로 버팁니다. 화면 동작에 맞춘 효과음과 스코어는 버티지 못하므로 두 번째 패스에서 만들고 `partially_copy`로 첫 트랙과 합치세요.
+
+### Model Sparse Attention으로 더 빠르게 샘플링하기 (#16072)
+
+[`Model Sparse Attention`](https://github.com/Comfy-Org/ComfyUI/pull/16072)은 전체 어텐션을 블록 희소 어텐션으로 대체합니다. 모델 라인에서 turbo LoRA 스위치(`ComfySwitchNode`)와 `BasicGuider` 사이에 모델 입력과 출력으로 연결하세요. ComfyUI **0.35.0+**, CUDA, 그리고 `sol-attn` 커널이 컴파일된 GPU가 필요합니다. 조건을 못 맞춘 호출은 밀집 어텐션으로 돌아가고 이유는 `verbose` 로그에 남습니다. 짧은 시퀀스는 밀집 어텐션이 더 빠른 경우가 많으므로 캔버스와 클립이 길수록 효과가 큽니다.
+
+| `method` | 필요한 가중치 |
+| --- | --- |
+| `sol-attn` | 없음. 어떤 H3 체크포인트에서도 동작합니다 |
+| `sla` | 이 선택 방식으로 증류된 모델. 베이스 H3 체크포인트로는 안 됩니다 |
+| `vsa` | [FastVideo 가속](/ko/tutorials/video/minimax/minimax-h3-fastvideo) 워크플로의 FastH3-VSA 체크포인트 |
+
+`start_percent`와 `end_percent`는 희소화가 적용되는 구간을 정합니다. 구간 밖은 밀집 어텐션으로 돌아갑니다. `start_percent` 기본값은 **0.2**, `end_percent`는 **1.0**이라 마지막 단계까지 희소 상태로 남습니다. 끝 지점을 **0.9**로 두면 마무리 단계가 전체 어텐션으로 디노이즈되고, 더 아끼려면 시작점을 0.2보다 앞당깁니다. 스케줄 중반은 양 끝보다 희소화를 더 잘 견딥니다.
+
+`tau`(기본 **1.3**)는 `sol-attn`의 속도와 품질 슬라이더입니다. 높을수록 더 희소해져 **1.0**은 키 블록의 약 16%, **1.5**는 약 7%, **2.0**은 약 2.7%를 정확히 계산합니다. 최적값은 없고, 이 맞교환 위의 한 지점을 고르는 문제입니다. `sla`와 `vsa`에서는 `tau` 대신 `keep_percent`(기본 **10**)를 쓰며, 각 쿼리 블록이 정확히 유지하는 키 블록 비율입니다.
+
+`min_tokens`(기본 **12288**)보다 짧은 시퀀스는 아무 표시 없이 밀집으로 실행됩니다. 16:9 템플릿 캔버스 **864×480**에서 5초 클립은 토큰이 약 **15k**라 희소로 돌고, 3초 클립은 **9k** 근처라 `min_tokens`를 낮추기 전까지 아무것도 줄지 않습니다. 이 수는 포워드 패스 한 번 기준이라 컨텍스트 윈도우를 연결하면 윈도우 단위가 되고, 해상도가 낮은 패스는 기준 아래로 내려갈 수 있습니다. `verbose`를 켜면 각 형상이 희소로 돌았는지, 밀집으로 남은 이유가 로그에 남습니다.
+
+`sink_conditioning`은 H3가 묶은 텍스트, 오디오, 참조 행을 정확하게 유지합니다. 기본값 `exact_kv_and_rows`는 생성된 오디오의 쿼리 행도 밀집으로 돌리므로, 영상 쪽을 희소화해도 사운드트랙은 그대로 유지됩니다.
diff --git a/tutorials/video/minimax/minimax-h3-native.mdx b/tutorials/video/minimax/minimax-h3-native.mdx
index 237e132aaa..c795e5fb9f 100644
--- a/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/tutorials/video/minimax/minimax-h3-native.mdx
@@ -264,11 +264,13 @@ Connect a mask to the sampler `denoise_mask`: **0** = keep, **1** = regenerate.
 
 Windows start on the model's 17-frame chunks, so video and audio stay in step across a boundary. `context_overlap` is capped at the window length minus **17** frames, and a larger value is clamped without a warning. A 5 s clip fits one window at the default **124**, and the first length that opens a second window is **141** frames. Lower `context_length` to cut the memory each step needs. Raise it for more context per step, at a higher cost.
 
+A long clip is cheaper as two passes. Render a low-resolution pass with long windows to settle motion and sound, then a higher-resolution pass with short windows to add detail. Window length trades against prompt following: each window is sampled as one shot, and the longer it runs, the sooner the clip stops following the prompt.
+
 Memory falls per step, not in total. The full latent, the buffers that fuse the windows and the decode still scale with clip length, so a clip that decodes out of memory stays out of memory.
 
 Two limits come with windowed sampling:
 
-- **Every window receives the full prompt.** Write one continuous scene and leave out timed cuts: a shot list with cut times assumes the whole clip is sampled in one pass.
+- **Every window receives the full prompt.** With windows on, write one continuous scene in place of a timed cut list: every window sees the whole prompt, so cut times no longer line up.
 - **Windows are fused, not joined.** Content that has to hold exactly across a boundary (on-screen text, set detail, a face) can drift. Keep the camera and the subject steady.
 
 That makes context windows a fit for a scene that runs without a cut: a slow push, a pan, or an unbroken take that runs past the window. For a clip that cuts between shots, render the shots separately and join them.
@@ -282,3 +284,21 @@ Audio and video are denoised together, so a short schedule or a turbo LoRA softe
 3. **Condition the second pass**: In the R2V workflow, add `LoadAudio`, wire it to `ref_audio`, and keep `audio_vae` connected. Cite the track as `<Audio 1>` and state the relationship: `reference` keeps the timbre, rhythm, and delivery without copying the signal, `fully_copy` makes it the complete final track, `partially_copy` keeps part of the timeline or selected layers, and `weak_reference` keeps only broad similarity. The second pass starts from a settled voice and timing instead of inventing speech from nothing. For the identical file, wire `LoadAudio` into `CreateVideo` instead of relying on the prompt.
 
 Canvas size trades voice quality for render time. A 32×32 pass is nearly free, but voices can pick up a robotic edge at that size. Raise the canvas when the voice has to hold up in the final cut, and expect the pass to cost more as you do. Speech and lyrics usually survive the small canvas. Sound effects and scoring tied to on-screen action do not, so render those in the second pass and let `partially_copy` combine them with the first track.
+
+### Faster sampling with Model Sparse Attention (#16072)
+
+[`Model Sparse Attention`](https://github.com/Comfy-Org/ComfyUI/pull/16072) runs block-sparse attention in place of full attention. Connect it on the model line, between the turbo LoRA switch (`ComfySwitchNode`) and `BasicGuider`, model in and model out. It needs ComfyUI **0.35.0+**, CUDA, and a GPU with a compiled `sol-attn` kernel. Otherwise every call falls back to dense attention and `verbose` logs the reason. Short sequences are usually faster dense, so the gain grows with canvas and clip length.
+
+| `method` | Weights it needs |
+| --- | --- |
+| `sol-attn` | None. Runs on any H3 checkpoint |
+| `sla` | A model distilled against this selection, not a base H3 checkpoint |
+| `vsa` | FastH3-VSA checkpoints from the [FastVideo speedup](/tutorials/video/minimax/minimax-h3-fastvideo) workflow |
+
+`start_percent` and `end_percent` bound where sparsification applies. Outside the window attention runs dense, and the node ships with `start_percent` **0.2** and `end_percent` **1.0**, which leaves the last steps sparse. End at **0.9** so the closing steps denoise at full attention, and start earlier than 0.2 when you want more savings: the middle of the schedule takes sparsification better than either end.
+
+`tau` (default **1.3**) is the speed and quality slider for `sol-attn`. Higher is sparser: **1.0** keeps about 16% of key blocks exact, **1.5** about 7%, **2.0** about 2.7%. There is no best value, only a point on that trade. `sla` and `vsa` swap `tau` for `keep_percent` (default **10**), the share of key blocks each query block keeps exactly.
+
+`min_tokens` (**12288**) leaves shorter sequences dense without saying so. At the 16:9 template canvas of **864×480**, a 5 s clip packs about **15k** tokens and runs sparse, while a 3 s clip stays near **9k** and gains nothing until you lower `min_tokens`. The count is per forward pass, so with context windows connected it is per window, and a low-resolution pass can fall under the gate. Set `verbose` to log whether each shape ran sparse and the reason it stayed dense.
+
+`sink_conditioning` keeps H3's packed text, audio and reference rows exact. The default `exact_kv_and_rows` also runs the generated-audio query rows dense, so the soundtrack holds up while the video attention is sparsified.
diff --git a/zh/tutorials/video/minimax/minimax-h3-native.mdx b/zh/tutorials/video/minimax/minimax-h3-native.mdx
index 87c0efc93d..76ad12c1ad 100644
--- a/zh/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/zh/tutorials/video/minimax/minimax-h3-native.mdx
@@ -2,14 +2,14 @@
 title: "ComfyUI MiniMax H3 文生视频、图生视频与参考生视频工作流"
 description: "ComfyUI 中 MiniMax H3 原生工作流的逐步示例，含帧锚定与潜在噪声蒙版技巧，帮助你对生成视频做更精细的控制。"
 sidebarTitle: "原生工作流"
-translationSourceHash: b16dd815
+translationSourceHash: 48763c0c
 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx
 translationBlockHashes:
   "_intro": dde651e5
   "MiniMax H3 Text to Video (T2V)": 42ec8cfb
   "MiniMax H3 Image to Video (I2V)": 2a06e3fa
   "MiniMax H3 Reference to Video (R2V)": 961829df
-  "Advanced workflows with the native nodes": d423a0db
+  "Advanced workflows with the native nodes": 2ab2df9d
 ---
 本页涵盖三个基础 MiniMax H3 模板及原生节点高级技巧。能力与授权见 [MiniMax H3 概览](/zh/tutorials/video/minimax/minimax-h3)。
 
@@ -271,11 +271,13 @@ ComfyUI/
 
 窗口从模型的 17 帧块边界开始，视频与音频跨边界仍对齐。`context_overlap` 上限是窗口长度减 **17** 帧，设得更大只会被静默压回上限。默认 **124** 下 5 秒片段占一个窗口，第一个能开出第二个窗口的长度是 **141** 帧。调小 `context_length` 降低每一步的显存占用。调大则每一步拿到更多上下文，代价更高。
 
+长片段用两遍更省。第一遍在低分辨率下用长窗口，先把运动与声音定下来；第二遍在高分辨率下用短窗口补细节。窗口长度与提示词跟随度是交换关系：每个窗口都当成一个镜头来采样，窗口越长，片段越早偏离提示词。
+
 省下的是每一步的显存，不是总量。整段 latent、融合窗口用的缓冲区和解码仍随片段长度增长，所以解码就会爆显存的片段，改完还是爆。
 
 窗口化采样自带两个限制：
 
-- **每个窗口都收到完整提示词。** 写一个连续的场景，不要写定时切镜：带切镜时间的分镜假定整段片段只采样一次。
+- **每个窗口都收到完整提示词。** 开了窗口后，用一段连续的场景替换带切镜时间的分镜：每个窗口看到的都是整段提示词，切镜时间对不上。
 - **窗口是融合，不是拼接。** 需要跨边界保持一致的内容（画面文字、场景细节、人脸）可能漂移。镜头和主体保持稳定。
 
 所以上下文窗口适合没有切镜的连续场景：缓慢推镜、平移，或超过窗口长度的一镜到底。需要切镜的片段，分镜单独生成再拼接。
@@ -289,3 +291,21 @@ ComfyUI/
 3. **给第二遍做条件**：在 R2V 工作流中加入 `LoadAudio`，接到 `ref_audio`，并保持 `audio_vae` 连接。提示词里用 `<Audio 1>` 引用它，并写明关系：`reference` 只取音色、节奏与语气，不复制信号本身；`fully_copy` 把它当作完整的最终音轨；`partially_copy` 只复制一部分时间线或部分音轨层；`weak_reference` 只保留大类与氛围上的相似。第二遍于是从一个已经确定的音色与节奏起步，而不是从零编造语音。要用完全相同的文件，把 `LoadAudio` 接到 `CreateVideo`，不要依赖提示词。
 
 画布尺寸是拿音色换渲染时间。32×32 的一遍几乎免费，但这个尺寸下声音可能带上机械感。音色要在成片里撑得住，就把画布调大，代价是这个遍数变贵。语音与歌词通常能扛住小画布；跟着画面动作走的音效与配乐扛不住，这类内容留给第二遍生成，再用 `partially_copy` 与第一遍的音轨合并。
+
+### 用 Model Sparse Attention 提速（#16072）
+
+[`Model Sparse Attention`](https://github.com/Comfy-Org/ComfyUI/pull/16072) 用块稀疏注意力替代全量注意力。它接在模型线上，位于 turbo LoRA 开关（`ComfySwitchNode`）与 `BasicGuider` 之间，模型进、模型出。需要 ComfyUI **0.35.0+**、CUDA，以及 `sol-attn` 内核编译覆盖到的显卡。不满足时每次调用都退回全量注意力，`verbose` 会把原因写进日志。短序列通常还是全量更快，所以画布与片段越长收益越大。
+
+| `method` | 需要的权重 |
+| --- | --- |
+| `sol-attn` | 不需要。任何 H3 检查点都能跑 |
+| `sla` | 针对该选择方式蒸馏过的模型，基础 H3 检查点不行 |
+| `vsa` | [FastVideo 提速](/zh/tutorials/video/minimax/minimax-h3-fastvideo)工作流里的 FastH3-VSA 检查点 |
+
+`start_percent` 与 `end_percent` 界定稀疏生效的区间，区间外回到全量注意力。节点默认 `start_percent` 为 **0.2**、`end_percent` 为 **1.0**，等于收尾几步也留在稀疏里。把结束点设成 **0.9**，让收尾阶段在全量注意力下收细；想更省就从 0.2 之前开始，调度中段的稀疏耐受度比两端好。
+
+`tau`（默认 **1.3**）是 `sol-attn` 的速度与质量滑杆。值越高越稀：**1.0** 大约保留 16% 的关键块精确计算，**1.5** 约 7%，**2.0** 约 2.7%。没有最优值，只有这条曲线上的某个点。`sla` 与 `vsa` 会把 `tau` 换成 `keep_percent`（默认 **10**），即每个查询块精确保留的关键块比例。
+
+`min_tokens`（默认 **12288**）会把更短的序列留在全量注意力，而且不提示。16:9 模板画布 **864×480** 下，5 秒片段约有 **15k** Token 走稀疏；3 秒片段只有 **9k** 左右，不调低 `min_tokens` 就一点没省。这个计数按单次前向算，接上上下文窗口后就是按窗口算，分辨率很低的那一遍可能直接落在阈值以下。打开 `verbose` 可以看到每次调用是否走了稀疏，以及没走的原因。
+
+`sink_conditioning` 让 H3 打包的文本、音频与参考行保持精确计算。默认的 `exact_kv_and_rows` 还会把生成音频的查询行按全量跑，所以视频侧稀疏化时音轨不受影响。

From fba7f2e6f266596977fe2eb46854b5d125db6fb9 Mon Sep 17 00:00:00 2001
From: lin-bot23 <linzhixiong23@gmail.com>
Date: Wed, 4 Nov 2026 18:07:12 +0800
Subject: [PATCH 13/13] docs(minimax): separate window length from clip length
 on H3 context windows

---
 es/tutorials/video/minimax/minimax-h3-native.mdx | 14 ++++++++------
 ja/tutorials/video/minimax/minimax-h3-native.mdx | 14 ++++++++------
 ko/tutorials/video/minimax/minimax-h3-native.mdx | 14 ++++++++------
 tutorials/video/minimax/minimax-h3-native.mdx    |  8 +++++---
 zh/tutorials/video/minimax/minimax-h3-native.mdx | 14 ++++++++------
 5 files changed, 37 insertions(+), 27 deletions(-)

diff --git a/es/tutorials/video/minimax/minimax-h3-native.mdx b/es/tutorials/video/minimax/minimax-h3-native.mdx
index f83718c07a..cc9e9b3944 100644
--- a/es/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/es/tutorials/video/minimax/minimax-h3-native.mdx
@@ -2,14 +2,14 @@
 title: "Flujos de trabajo de MiniMax H3 en ComfyUI: texto a video, imagen a video y referencia a video"
 description: "Ejemplos paso a paso de flujos nativos de MiniMax H3 en ComfyUI, con anclaje de guías y máscaras de ruido latente para un control más fino del video."
 sidebarTitle: "Flujos nativos"
-translationSourceHash: 48763c0c
+translationSourceHash: 2d2c970c
 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx
 translationBlockHashes:
   "_intro": dde651e5
-  "MiniMax H3 Text to Video (T2V)": 42ec8cfb
+  "MiniMax H3 Text to Video (T2V)": 0cbf3427
   "MiniMax H3 Image to Video (I2V)": 2a06e3fa
   "MiniMax H3 Reference to Video (R2V)": 961829df
-  "Advanced workflows with the native nodes": 2ab2df9d
+  "Advanced workflows with the native nodes": 0a482892
 ---
 Esta página cubre las tres plantillas base de MiniMax H3 y las técnicas avanzadas con nodos nativos. Para conocer las capacidades y el licenciamiento, consulta la [descripción general de MiniMax H3](/es/tutorials/video/minimax/minimax-h3).
 
@@ -43,7 +43,7 @@ Genera videos a partir de prompts de texto con audio estéreo nativo.
   <Card title="VAE: minimax_h3_audio_vae_fp32" icon="download" href="https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_audio_vae_fp32.safetensors">
     Coloca en <code>ComfyUI/models/vae/</code>
   </Card>
-  <Card title="LoRA: minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16" icon="download" href="https://huggingface.co/lightx2v/Minimax-h3-Turbo/resolve/main/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors">
+  <Card title="LoRA: minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16" icon="download" href="https://huggingface.co/lightx2v/Minimax-h3-Turbo/blob/main/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors">
     Requerido por el escaneo de modelos del flujo de trabajo; habilita el modo turbo opcional de Lightning LoRA. Coloca en <code>ComfyUI/models/loras/</code>
   </Card>
 </CardGroup>
@@ -271,6 +271,8 @@ Conecta una máscara a `denoise_mask` del sampler: **0** = mantener, **1** = reg
 
 Las ventanas empiezan en los bloques de 17 frames del modelo, de modo que video y audio siguen alineados al cruzar un límite. `context_overlap` tiene como tope la longitud de la ventana menos **17** frames, y un valor mayor se recorta sin aviso. Con el valor por defecto **124** un clip de 5 s ocupa una sola ventana, y la primera longitud que abre una segunda ventana es **141** frames. Baja `context_length` para reducir la memoria de cada paso. Súbelo para dar más contexto a cada paso, con un coste mayor.
 
+`context_length` limita cuánto del clip ve cada paso de muestreo, no la duración del clip. La duración la fija la entrada `length` (un conteo de frames a 24 fps) del nodo H3, así que un clip más largo que una ventana se muestrea igualmente de principio a fin, ventana a ventana, y conserva la duración que fijaste. El rango entrenado es de unos **124** a **362** frames a 24 fps (**5** a **15** s); los clips más largos no están probados.
+
 Un clip largo sale más barato en dos pasadas. Renderiza una pasada a baja resolución con ventanas largas para asentar movimiento y sonido, y después otra a mayor resolución con ventanas cortas para añadir detalle. La longitud de ventana se cambia por fidelidad al prompt: cada ventana se muestrea como un solo plano, y cuanto más larga es, antes deja el clip de seguir el prompt.
 
 La memoria baja por paso, no en total. El latente completo, los búferes que fusionan las ventanas y el decodificado siguen creciendo con la longitud del clip, así que un clip que se queda sin memoria al decodificar sigue igual.
@@ -278,9 +280,9 @@ La memoria baja por paso, no en total. El latente completo, los búferes que fus
 El muestreo por ventanas trae dos límites:
 
 - **Cada ventana recibe el prompt completo.** Con ventanas activas, describe una escena continua en lugar de un desglose de planos con marcas de tiempo: todas las ventanas ven el prompt entero, así que los tiempos de corte ya no encajan.
-- **Las ventanas se fusionan, no se empalman.** El contenido que debe mantenerse idéntico al cruzar un límite (texto en pantalla, detalle del decorado, una cara) puede derivar. Mantén estables la cámara y el sujeto.
+- **Las ventanas se fusionan, no se empalman.** El contenido que debe mantenerse idéntico al cruzar un límite (texto en pantalla, detalle del decorado, una cara) puede derivar.
 
-Por eso las ventanas de contexto encajan en una escena sin cortes: un push lento, un paneo o un plano sin corte que supera la ventana. Para un clip que cambia de plano, genera los planos por separado y únelos.
+Por eso las ventanas de contexto encajan en una toma larga de una escena que se mantiene quieta: cámara fija, sujeto estable y un fondo que no cambia mucho. Si la cámara recorre un decorado o el sujeto camina delante del fondo, cada ventana ve ese fondo de otra manera y las uniones se ven como deformación. Para un clip así, genera tramos más cortos y únelos.
 
 ### Generar primero la banda sonora
 
diff --git a/ja/tutorials/video/minimax/minimax-h3-native.mdx b/ja/tutorials/video/minimax/minimax-h3-native.mdx
index 9eb00b8b29..39cd4e7ba7 100644
--- a/ja/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/ja/tutorials/video/minimax/minimax-h3-native.mdx
@@ -2,14 +2,14 @@
 title: "ComfyUI MiniMax H3 テキストから動画・画像から動画・参照から動画のワークフロー"
 description: "ComfyUI における MiniMax H3 ネイティブワークフローの手順例。ガイドのアンカー固定や latent ノイズマスクで生成動画をより細かく制御できます。"
 sidebarTitle: "ネイティブワークフロー"
-translationSourceHash: 48763c0c
+translationSourceHash: 2d2c970c
 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx
 translationBlockHashes:
   "_intro": dde651e5
-  "MiniMax H3 Text to Video (T2V)": 42ec8cfb
+  "MiniMax H3 Text to Video (T2V)": 0cbf3427
   "MiniMax H3 Image to Video (I2V)": 2a06e3fa
   "MiniMax H3 Reference to Video (R2V)": 961829df
-  "Advanced workflows with the native nodes": 2ab2df9d
+  "Advanced workflows with the native nodes": 0a482892
 ---
 
 このページは 3 つの基本 MiniMax H3 テンプレートとネイティブノードの高度なテクニックを扱います。機能とライセンスは [MiniMax H3 概要](/ja/tutorials/video/minimax/minimax-h3) を参照してください。
@@ -44,7 +44,7 @@ translationBlockHashes:
   <Card title="VAE: minimax_h3_audio_vae_fp32" icon="download" href="https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_audio_vae_fp32.safetensors">
     <code>ComfyUI/models/vae/</code> に配置します
   </Card>
-  <Card title="LoRA: minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16" icon="download" href="https://huggingface.co/lightx2v/Minimax-h3-Turbo/resolve/main/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors">
+  <Card title="LoRA: minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16" icon="download" href="https://huggingface.co/lightx2v/Minimax-h3-Turbo/blob/main/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors">
     ワークフローのモデルスキャンに必須。オプションの Lightning LoRA turbo 用。<code>ComfyUI/models/loras/</code> に配置します
   </Card>
 </CardGroup>
@@ -272,6 +272,8 @@ H3 ノードの `positive` と latent を `MiniMaxH3AddGuide` に接続:
 
 ウィンドウはモデルの 17 フレーム塊の境界から始まるため、動画と音声は境界をまたいでも揃います。`context_overlap` の上限はウィンドウ長から **17** フレームを引いた値で、それより大きく設定しても警告なしに切り詰められます。デフォルトの **124** では 5 秒のクリップが 1 ウィンドウに収まり、2 つ目のウィンドウが開く最小の長さは **141** フレームです。`context_length` を下げると 1 ステップあたりのメモリが減ります。上げると 1 ステップあたりの前後文脈が増え、コストも上がります。
 
+`context_length` が決めるのは各ステップで見る範囲で、クリップ自体の長さではありません。クリップ長は H3 ノードの `length` 入力（24fps でのフレーム数）で決まるため、ウィンドウより長いクリップも 1 ウィンドウずつ最後までサンプリングされ、指定した長さのまま出力されます。学習範囲は約 **124** から **362** フレーム（24fps で **5** から **15** 秒）で、それより長いクリップは未検証です。
+
 長いクリップは 2 パスに分けるほうが安く済みます。低解像度で長いウィンドウを使い、まず動きと音を固めます。そのあと高解像度で短いウィンドウを使い、ディテールを足します。ウィンドウ長とプロンプト追従は交換関係にあり、各ウィンドウは 1 ショットとしてサンプリングされるため、長くするほどクリップがプロンプトから外れるのが早まります。
 
 減るのは 1 ステップあたりのメモリで、総量ではありません。全体の latent、ウィンドウ融合用のバッファ、デコードはクリップ長に比例したままなので、デコードで OOM になるクリップはそのまま OOM になります。
@@ -279,9 +281,9 @@ H3 ノードの `positive` と latent を `MiniMaxH3AddGuide` に接続:
 ウィンドウ分割には 2 つの制約があります:
 
 - **どのウィンドウにもプロンプト全体が渡ります。** ウィンドウを使うときは、時刻付きのカット割りではなく 1 つの連続したシーンとして書いてください。どのウィンドウもプロンプト全体を見るため、カット時刻は噛み合いません。
-- **ウィンドウは融合されるのであって連結ではありません。** 境界をまたいで厳密に保つ必要がある内容（画面内の文字、セットの細部、顔）はドリフトし得ます。カメラと被写体は安定させます。
+- **ウィンドウは融合されるのであって連結ではありません。** 境界をまたいで厳密に保つ必要がある内容（画面内の文字、セットの細部、顔）はドリフトし得ます。
 
-そのためコンテキストウィンドウは、カットのない連続したシーン（ゆっくりしたプッシュイン、パン、ウィンドウ長を超えるワンカット）に向きます。ショットを切り替えるクリップは、ショットごとに生成してからつなぎます。
+そのためコンテキストウィンドウは、場面が動かない長回しに向きます。固定カメラ、安定した被写体、大きく変わらない背景です。カメラがセットの中を移動したり、被写体が背景の前を歩いたりすると、各ウィンドウで背景の見え方が変わり、継ぎ目が変形として出ます。そうしたクリップは短く分けて生成し、つないでください。
 
 ### 音声を先に作る
 
diff --git a/ko/tutorials/video/minimax/minimax-h3-native.mdx b/ko/tutorials/video/minimax/minimax-h3-native.mdx
index bfd8bb8c9d..9b53eac726 100644
--- a/ko/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/ko/tutorials/video/minimax/minimax-h3-native.mdx
@@ -2,14 +2,14 @@
 title: "ComfyUI MiniMax H3 텍스트·이미지·레퍼런스 기반 비디오 워크플로"
 description: "ComfyUI의 MiniMax H3 네이티브 워크플로 단계별 예제입니다. 가이드 고정과 잠재 노이즈 마스크로 생성 비디오를 더 세밀하게 제어할 수 있습니다."
 sidebarTitle: "네이티브 워크플로"
-translationSourceHash: 48763c0c
+translationSourceHash: 2d2c970c
 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx
 translationBlockHashes:
   "_intro": dde651e5
-  "MiniMax H3 Text to Video (T2V)": 42ec8cfb
+  "MiniMax H3 Text to Video (T2V)": 0cbf3427
   "MiniMax H3 Image to Video (I2V)": 2a06e3fa
   "MiniMax H3 Reference to Video (R2V)": 961829df
-  "Advanced workflows with the native nodes": 2ab2df9d
+  "Advanced workflows with the native nodes": 0a482892
 ---
 
 이 페이지는 세 가지 기본 MiniMax H3 템플릿과 네이티브 노드 고급 기법을 다룹니다. 기능과 라이선스는 [MiniMax H3 개요](/ko/tutorials/video/minimax/minimax-h3)를 참조하세요.
@@ -44,7 +44,7 @@ translationBlockHashes:
   <Card title="VAE: minimax_h3_audio_vae_fp32" icon="download" href="https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_audio_vae_fp32.safetensors">
     <code>ComfyUI/models/vae/</code> 폴더에 넣으세요.
   </Card>
-  <Card title="LoRA: minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16" icon="download" href="https://huggingface.co/lightx2v/Minimax-h3-Turbo/resolve/main/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors">
+  <Card title="LoRA: minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16" icon="download" href="https://huggingface.co/lightx2v/Minimax-h3-Turbo/blob/main/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors">
     워크플로 모델 스캔에 필요; 선택 Lightning LoRA turbo. <code>ComfyUI/models/loras/</code>에 넣으세요.
   </Card>
 </CardGroup>
@@ -272,6 +272,8 @@ H3 노드 `positive`·latent를 `MiniMaxH3AddGuide`에 연결:
 
 윈도우는 모델의 17프레임 덩어리 경계에서 시작하므로 영상과 오디오가 경계를 넘어도 어긋나지 않습니다. `context_overlap` 상한은 윈도우 길이에서 **17**프레임을 뺀 값이고, 더 크게 넣으면 경고 없이 잘립니다. 기본값 **124**에서는 5초 클립이 윈도우 하나에 들어가고, 두 번째 윈도우가 열리는 최소 길이는 **141**프레임입니다. `context_length`를 낮추면 단계마다 쓰는 메모리가 줄어듭니다. 높이면 단계마다 더 많은 앞뒤 맥락을 보는 대신 비용이 커집니다.
 
+`context_length`는 각 단계가 보는 범위를 정할 뿐 클립 자체의 길이를 정하지 않습니다. 클립 길이는 H3 노드의 `length` 입력(24fps 기준 프레임 수)이라, 윈도우보다 긴 클립도 윈도우 하나씩 끝까지 샘플링되고 지정한 길이대로 출력됩니다. 학습 범위는 약 **124**에서 **362**프레임(24fps에서 **5**에서 **15**초)이고, 더 긴 클립은 검증되지 않았습니다.
+
 긴 클립은 두 패스로 나누는 편이 저렴합니다. 낮은 해상도에서 긴 윈도우로 먼저 움직임과 소리를 잡고, 그다음 높은 해상도에서 짧은 윈도우로 세부를 더합니다. 윈도우 길이는 프롬프트 준수와 맞바꾸는 값이라, 각 윈도우가 한 숏으로 샘플링되므로 길수록 클립이 프롬프트에서 일찍 벗어납니다.
 
 줄어드는 것은 단계당 메모리이지 총량이 아닙니다. 전체 latent, 윈도우를 합치는 버퍼, 디코드는 클립 길이에 비례하므로, 디코드에서 메모리가 부족한 클립은 그대로 부족합니다.
@@ -279,9 +281,9 @@ H3 노드 `positive`·latent를 `MiniMaxH3AddGuide`에 연결:
 윈도우 분할에는 제약이 두 가지 있습니다:
 
 - **모든 윈도우가 프롬프트 전체를 받습니다.** 윈도우를 켜면 시각이 붙은 컷 구성 대신 하나의 연속된 장면으로 쓰세요. 모든 윈도우가 프롬프트 전체를 보므로 컷 시각이 맞지 않습니다.
-- **윈도우는 이어 붙이는 게 아니라 합쳐집니다.** 경계를 넘어 정확히 유지해야 하는 내용(화면 속 글자, 세트 디테일, 얼굴)은 밀릴 수 있습니다. 카메라와 피사체는 안정적으로 유지합니다.
+- **윈도우는 이어 붙이는 게 아니라 합쳐집니다.** 경계를 넘어 정확히 유지해야 하는 내용(화면 속 글자, 세트 디테일, 얼굴)은 밀릴 수 있습니다.
 
-그래서 컨텍스트 윈도우는 컷이 없는 연속 장면에 맞습니다. 느린 푸시인, 팬, 윈도우보다 긴 원테이크 같은 경우입니다. 샷을 전환하는 클립은 샷별로 생성한 뒤 이어 붙입니다.
+그래서 컨텍스트 윈도우는 장면이 거의 움직이지 않는 긴 테이크에 맞습니다. 고정 카메라, 안정적인 피사체, 크게 변하지 않는 배경입니다. 카메라가 세트를 지나가거나 피사체가 배경 앞을 걸으면 윈도우마다 배경이 다르게 보이게 되고, 이음새가 변형으로 드러납니다. 그런 클립은 짧게 나눠 생성한 뒤 이어 붙입니다.
 
 ### 사운드트랙을 먼저 만들기
 
diff --git a/tutorials/video/minimax/minimax-h3-native.mdx b/tutorials/video/minimax/minimax-h3-native.mdx
index c795e5fb9f..c6204eb3de 100644
--- a/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/tutorials/video/minimax/minimax-h3-native.mdx
@@ -36,7 +36,7 @@ Generate videos from text prompts with native stereo audio.
   <Card title="VAE: minimax_h3_audio_vae_fp32" icon="download" href="https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_audio_vae_fp32.safetensors">
     Place in <code>ComfyUI/models/vae/</code>
   </Card>
-  <Card title="LoRA: minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16" icon="download" href="https://huggingface.co/lightx2v/Minimax-h3-Turbo/resolve/main/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors">
+  <Card title="LoRA: minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16" icon="download" href="https://huggingface.co/lightx2v/Minimax-h3-Turbo/blob/main/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors">
     Required by the workflow's model scan; powers the optional Lightning LoRA turbo mode. Place in <code>ComfyUI/models/loras/</code>
   </Card>
 </CardGroup>
@@ -264,6 +264,8 @@ Connect a mask to the sampler `denoise_mask`: **0** = keep, **1** = regenerate.
 
 Windows start on the model's 17-frame chunks, so video and audio stay in step across a boundary. `context_overlap` is capped at the window length minus **17** frames, and a larger value is clamped without a warning. A 5 s clip fits one window at the default **124**, and the first length that opens a second window is **141** frames. Lower `context_length` to cut the memory each step needs. Raise it for more context per step, at a higher cost.
 
+`context_length` bounds how much of the clip each sampling step sees, not how long the clip is. Clip length is the `length` input (a frame count at 24 fps) on the H3 node, so a clip longer than one window is still sampled end to end, one window at a time, and keeps the duration you set. The trained range is about **124** to **362** frames at 24 fps (**5** to **15** s); longer clips are untested.
+
 A long clip is cheaper as two passes. Render a low-resolution pass with long windows to settle motion and sound, then a higher-resolution pass with short windows to add detail. Window length trades against prompt following: each window is sampled as one shot, and the longer it runs, the sooner the clip stops following the prompt.
 
 Memory falls per step, not in total. The full latent, the buffers that fuse the windows and the decode still scale with clip length, so a clip that decodes out of memory stays out of memory.
@@ -271,9 +273,9 @@ Memory falls per step, not in total. The full latent, the buffers that fuse the
 Two limits come with windowed sampling:
 
 - **Every window receives the full prompt.** With windows on, write one continuous scene in place of a timed cut list: every window sees the whole prompt, so cut times no longer line up.
-- **Windows are fused, not joined.** Content that has to hold exactly across a boundary (on-screen text, set detail, a face) can drift. Keep the camera and the subject steady.
+- **Windows are fused, not joined.** Content that has to hold exactly across a boundary (on-screen text, set detail, a face) can drift.
 
-That makes context windows a fit for a scene that runs without a cut: a slow push, a pan, or an unbroken take that runs past the window. For a clip that cuts between shots, render the shots separately and join them.
+That makes context windows a fit for a long take of a scene that holds still: a locked-off camera, a steady subject, a background that does not change much. A camera move through a set, or a subject walking past a background, gives each window a different view of that background, and the joins show as morphing. For a clip like that, render shorter clips and join them.
 
 ### Generating the soundtrack first
 
diff --git a/zh/tutorials/video/minimax/minimax-h3-native.mdx b/zh/tutorials/video/minimax/minimax-h3-native.mdx
index 76ad12c1ad..ca39d3813e 100644
--- a/zh/tutorials/video/minimax/minimax-h3-native.mdx
+++ b/zh/tutorials/video/minimax/minimax-h3-native.mdx
@@ -2,14 +2,14 @@
 title: "ComfyUI MiniMax H3 文生视频、图生视频与参考生视频工作流"
 description: "ComfyUI 中 MiniMax H3 原生工作流的逐步示例，含帧锚定与潜在噪声蒙版技巧，帮助你对生成视频做更精细的控制。"
 sidebarTitle: "原生工作流"
-translationSourceHash: 48763c0c
+translationSourceHash: 2d2c970c
 translationFrom: tutorials/video/minimax/minimax-h3-native.mdx
 translationBlockHashes:
   "_intro": dde651e5
-  "MiniMax H3 Text to Video (T2V)": 42ec8cfb
+  "MiniMax H3 Text to Video (T2V)": 0cbf3427
   "MiniMax H3 Image to Video (I2V)": 2a06e3fa
   "MiniMax H3 Reference to Video (R2V)": 961829df
-  "Advanced workflows with the native nodes": 2ab2df9d
+  "Advanced workflows with the native nodes": 0a482892
 ---
 本页涵盖三个基础 MiniMax H3 模板及原生节点高级技巧。能力与授权见 [MiniMax H3 概览](/zh/tutorials/video/minimax/minimax-h3)。
 
@@ -43,7 +43,7 @@ translationBlockHashes:
   <Card title="VAE：minimax_h3_audio_vae_fp32" icon="download" href="https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/vae/minimax_h3_audio_vae_fp32.safetensors">
     放入 <code>ComfyUI/models/vae/</code>
   </Card>
-  <Card title="LoRA：minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16" icon="download" href="https://huggingface.co/lightx2v/Minimax-h3-Turbo/resolve/main/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors">
+  <Card title="LoRA：minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16" icon="download" href="https://huggingface.co/lightx2v/Minimax-h3-Turbo/blob/main/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors">
     工作流的模型扫描需要此文件；它为可选的 Lightning LoRA 加速模式提供支持。放入 <code>ComfyUI/models/loras/</code>
   </Card>
 </CardGroup>
@@ -271,6 +271,8 @@ ComfyUI/
 
 窗口从模型的 17 帧块边界开始，视频与音频跨边界仍对齐。`context_overlap` 上限是窗口长度减 **17** 帧，设得更大只会被静默压回上限。默认 **124** 下 5 秒片段占一个窗口，第一个能开出第二个窗口的长度是 **141** 帧。调小 `context_length` 降低每一步的显存占用。调大则每一步拿到更多上下文，代价更高。
 
+`context_length` 限制的是每一步采样能看到的片段长度，不是片段本身多长。片段长度由 H3 节点的 `length` 输入（24fps 下的帧数）决定，所以超过一个窗口的片段仍会按你设定的时长整段采样，一次一个窗口。模型训练覆盖约 **124** 到 **362** 帧（24fps 下 **5** 到 **15** 秒），更长的片段未经验证。
+
 长片段用两遍更省。第一遍在低分辨率下用长窗口，先把运动与声音定下来；第二遍在高分辨率下用短窗口补细节。窗口长度与提示词跟随度是交换关系：每个窗口都当成一个镜头来采样，窗口越长，片段越早偏离提示词。
 
 省下的是每一步的显存，不是总量。整段 latent、融合窗口用的缓冲区和解码仍随片段长度增长，所以解码就会爆显存的片段，改完还是爆。
@@ -278,9 +280,9 @@ ComfyUI/
 窗口化采样自带两个限制：
 
 - **每个窗口都收到完整提示词。** 开了窗口后，用一段连续的场景替换带切镜时间的分镜：每个窗口看到的都是整段提示词，切镜时间对不上。
-- **窗口是融合，不是拼接。** 需要跨边界保持一致的内容（画面文字、场景细节、人脸）可能漂移。镜头和主体保持稳定。
+- **窗口是融合，不是拼接。** 需要跨边界保持一致的内容（画面文字、场景细节、人脸）可能漂移。
 
-所以上下文窗口适合没有切镜的连续场景：缓慢推镜、平移，或超过窗口长度的一镜到底。需要切镜的片段，分镜单独生成再拼接。
+所以上下文窗口适合场景基本静止的长镜头：固定机位、主体稳定、背景变化不大。镜头穿过场景，或主体走过背景，都会让每个窗口看到不同的背景，衔接处表现为形变。这类片段请分段生成再拼接。
 
 ### 先生成音轨