Imported from Hello-Hoy/CatVTON_test (.codex/skills/catvton-train-reconstruction/SKILL.md). Install upstream with npx skills add Hello-Hoy/CatVTON_test --skill catvton-train-reconstruction. Copyright stays with the author.

CatVTON Train Reconstruction

Use this skill for work inside CatVTON_practice.

What this repo is for

This project is a practice implementation of CatVTON-style training centered on the local DressCode dataset under data/DressCode.

The important local constraint is that DressCode here does not already contain agnostic_masks, so training must either generate them on the fly or cache them before use.

Workflow

1. Read references/repo-notes.md.
2. Keep the implementation scoped to this repo before reaching back into the original CatVTON repo.
3. Preserve CatVTON-compatible checkpoint layout:
   • <output>/<dataset_tag>/attention
4. Prefer self-attention-only training unless the user explicitly asks for a different fine-tuning target.
5. When changing dataset logic, validate against the actual files under data/DressCode.

Validation rules

• Confirm train_pairs.txt and test pair files still parse.
• Confirm masks are generated or loaded consistently.
• Confirm the saved checkpoint contains attention weights that can be reloaded later.

References

• references/repo-notes.md