This folder contains research notebooks used to compare models, tune pipeline settings, and analyze benchmark outputs.
Notebooks are for exploration and thesis evidence preparation. The source-of-truth workflows stay in code, manifests, benchmark outputs, and folder-level READMEs. When a notebook produces a final claim, save the supporting table or figure into thesis/ and reference the benchmark files that produced it.
| Path | Focus |
|---|---|
comparison/ |
Model-to-model comparisons and selective-vs-full processing studies |
frame-selection/ |
Scene-change, blur, and sparse sampling experiments |
optimization/ |
Batch size, quantization, and runtime tuning |
quality-analysis/ |
OCR, PSNR/SSIM, and subjective quality analysis |
| Question | Likely evidence |
|---|---|
| Does EduScale improve readability compared with LR input or bicubic upscaling? | OCR confidence, CER, and visual comparison notebooks |
| Which model checkpoint is best for x2 and x3? | Benchmark summary tables and model comparison notebooks |
| Does OCR-aware training improve text recognition? | quality-analysis/ocr-accuracy-analysis.ipynb and OCR summary JSON files |
| What is the runtime tradeoff of model size and quantization? | optimization/quantization-experiments.ipynb and runtime benchmarks |
| Can frame selection reduce unnecessary processing? | frame-selection/ notebooks |
- Generate benchmark outputs in
benchmarks/results/oroutputs/benchmarks/. - Open the notebook in VS Code or Jupyter.
- Update input paths only when analyzing a new run.
- Save final plots into
thesis/figures/. - Save final tables into thesis-ready CSV or Markdown.
- Record the benchmark file names behind any reported number.
- exploratory analysis
- ablation comparisons
- visualization notebooks
- thesis figure preparation
- runtime and quality tradeoff analysis
- production training code
- canonical benchmark scripts
- raw datasets
- large generated images or videos
- long-term workflow documentation