docs(minimax, ltx, flux): H3 sparse attention, canvas size, camera motion, dialogue rules and context windows, LTX-2.5 VAE choice, FLUX 3 Klein layout path - #2121

Adds usage sections across four pages.

MiniMax H3 prompt guide: a ### Camera motion section with the motion vocabulary from the official base prompt guide, written into the shot as an action with amplitude and speed.

MiniMax H3 prompt guide, dialogue rules: three bullets in ## Dialogue and speakers now cover where a line may appear and how it ends. retention_analysis takes no dialogue, lyrics, or (Sx) IDs, and each spoken line stays in <d> in the shot where it is heard. An <Audio N> used only for timbre, rhythm, emotion, or delivery no longer carries its words into the target. Each line closes with one terminal mark before </d>, with no repeated marks and nothing after the tag.

MiniMax H3 prompt guide, speaker identity: ## Dialogue and speakers now states that IDs index speakers and hold across shots, that characters who never vocalize get none, and what to write outside the tag on first appearance: character type, age, gender, on-screen or off-screen, pitch, timbre, speaking rate, accent. The accent goes in the identifying phrase, not the [English] tag. <cutoff> covers a line the end of the clip truncates.

MiniMax H3 prompt guide, canvas size: a ## Canvas size section covering the canvas the templates actually open at. All three ship 0.4 MP, 864×480 at 16:9 for T2V and R2V and 640×640 at 1:1 for I2V, while the 16:9 native canvas is 1344×768, listed at 0.98 MP in each template's Size Settings Reference note. Canvas size changes the video latent, the audio latent follows clip length only, and both are generated in the same pass, so the same prompt and seed give different dialogue at a different canvas. R2V also downscales reference images toward the canvas area, so a larger canvas keeps more reference detail.

MiniMax H3 native workflows: two turbo notes (fast action and camera moves soften first under a short turbo schedule) and a ### Generating the soundtrack first section. The audio pass runs at 32×32 with Boolean (Enable Lightning LoRA) off and 40–50 steps, saves through SaveAudioAdvanced, then feeds ref_audio in R2V with an explicit relationship marker. Audio latent frames follow clip length, not canvas size, so the cheap pass still produces a full-length track.

MiniMax H3 native workflows, context windows: a ### Long clips with context windows section for the MiniMaxH3ContextWindows node. It covers where the node sits on the model line, the 17n+5 grid on context_length and context_overlap, the overlap cap, why the default 124 only changes sampling above 141 frames, and what windowing does not fix: total memory, timed cuts, and drift across a window boundary. The same pass corrects the soundtrack step note, which named the internal 6 as the turbo step count while the subgraph widget sets 8.

LTX-2.5: a Note on the two video VAEs in the pack. The templates load the diffusion decoder, which cannot be offloaded, so a full-size decode holds a lot of VRAM; ltx-2.5-video-vae-conv-bf16 decodes through the convolutional autoencoder when that decode runs out of memory.

FLUX 3 Klein: the single Create Bounding Boxes bullet becomes six describing how the boxes reach the model. Create Bounding Boxes puts them on its elements output and Flux3LayoutPrompt appends them to the prompt as a JSON layout table, cited as <obj_1>. The image edit templates route that table through ComfySwitchNode, which ships on the upsampler branch where the text is rewritten and the table lost, so the page says to set the switch to true. The two text to image templates leave the layout output on Preview Any, so it has to be dragged onto the subgraph prompt input with enable_prompt_upsampler off. The bullets also cover covering the whole canvas with one box for text to image, the box canvas aspect ratio, and colours coming from the box description rather than the editor palette.

Two corrections came out of that pass. The image edit step said to size the output with Empty Flux 2 Latent, which is not connected in either template: the output follows the reference through ReferenceMultiLatent.empty_latent. The upsampler download card said the file was used by the text to image templates, while all four load it.

The es pages also regain three node names that had been translated: ModelSamplingMiniMaxH3, Model Attention Backend, Model Sparse Attention.

Basis

Camera-motion values come from the official base prompt guide section 4.3. The audio pass follows temporal_shape() and _encode_ref_audio() in comfy_extras/nodes_minimax_h3.py, the 0.1 Megapixels floor in comfy_extras/nodes_resolution.py, and the audio relationship markers in the official ref guide section 4.2. The 20-step PrimitiveInt, the turbo count the subgraph widget sets to 8, and the Boolean (Enable Lightning LoRA) toggle are read from video_minimax_h3_t2v.json.

Canvas figures follow temporal_shape() and _empty_av_latent() in comfy_extras/nodes_minimax_h3.py, where the video latent is [B, 24, T, H/16, W/16] and the audio latent is [B, 32, 2, round(duration × 40)] in one nested tensor, plus the ResolutionSelector maths in comfy_extras/nodes_resolution.py and the ResolutionSelector widgets and Note: Size Settings Reference note in video_minimax_h3_t2v.json, video_minimax_h3_i2v.json and video_minimax_h3_r2v.json.

The speaker rules follow section 4.4 of the official base guide: identity attributes are written outside <d>, a speaker keeps one ID across shots, characters who never vocalize carry no ID, and <cutoff> marks speech truncated by the video end.

The dialogue rules follow section 5.4 of the official ref guide: the terminal mark goes before </d>, (Sx) is never written in retention_analysis, and a reference used only for timbre or delivery does not bring its words along. The reference marker example is the guide's own line from section 4.2.

The two VAEs and their tradeoff follow the LTX-2.5 model card component table and the two decoder branches in comfy/sd.py, which also sets disable_offload for the diffusion decoder.

Context-window behaviour follows MiniMaxH3ContextWindowsNode in comfy_extras/nodes_context_windows.py (17n+5 steps on both inputs, overlap clamped to the window in latent frames) and video_latent_t() in comfy_extras/nodes_minimax_h3.py; the full-length accumulators that keep the decode and the fuse buffers at clip length are in comfy/context_windows.py. The wiring follows BasicGuider.model and BasicScheduler.model in video_minimax_h3_t2v.json, where both take the ComfySwitchNode output.

The layout path follows Flux3LayoutPrompt in comfy_extras/nodes_flux.py, CreateBoundingBoxes and ReferenceMultiLatent in comfy_extras/nodes_bounding_boxes.py and comfy_extras/nodes_edit_model.py, and SwitchNode in comfy_extras/nodes_logic.py, checked against the four templates in Comfy-Org/workflow_templates.

zh, ja, ko, and es are translated in the same commit.

MiniMax H3 native workflows, sparse attention: a ### Faster sampling with Model Sparse Attention (#16072) section. It covers where the node sits on the model line, the weights each method needs, the start_percent / end_percent window (0.2 to 1.0 by default) and why ending at 0.9 returns the closing steps to dense attention, tau (default 1.3) and keep_percent (default 10) as the speed and quality trade, and the min_tokens gate (12288) that leaves shorter sequences dense without a message, with the counts a 5 s and a 3 s clip reach at the 864×480 template canvas. The count is per forward pass, so it is per window once context windows are connected. Off CUDA, or without a compiled kernel, every call falls back to dense attention.

MiniMax H3 native workflows, context windows: one paragraph on running a long clip as two passes, low resolution with long windows then higher resolution with short windows, and the window-length trade against prompt following. The bullet on the full prompt now says what to do with a timed cut list when windows are on.

Basis: the BlockSparseAttention schema, dense_reason(), _ineligible() and sinks() in comfy_extras/nodes_sparse_attention.py; temporal_shape(), video_latent_t() and align_frame_count() in comfy_extras/nodes_minimax_h3.py; ResolutionSelector in comfy_extras/nodes_resolution.py; and the ComfySwitchNode to BasicGuider wiring in video_minimax_h3_t2v.json. The vsa path links the FastVideo page.

MiniMax H3 native workflows, context windows (follow-up): the section now separates window length from clip length. context_length bounds what each sampling step sees, while the length input on the H3 node sets the clip duration, so a clip longer than one window is still sampled end to end. The trained range is about 124 to 362 frames at 24 fps, and longer clips are untested.

The section also names the scene windowing holds: a locked-off camera, a steady subject, and 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. Those clips are better rendered shorter and joined.

Basis: the length input and its tooltip in comfy_extras/nodes_minimax_h3.py, and MiniMaxH3ContextWindowsNode.execute(), which caps the overlap at context_length - 5 latent frames.

The T2V turbo LoRA card now uses the blob/main URL the rest of the page uses.