catvton-train-reconstruction
Use when implementing, reviewing, or extending CatVTON-style training in this repository. Focus on DressCode-based training, agnostic mask generation from local dataset annotations, and attention checkpoint export compatible with CatVTON inference layouts.
Install
mkdir -p .claude/skills/catvton-train-reconstruction-hello-hoy && curl -L -o skill.zip "https://agentskills.codes/api/skills/download/14777" && unzip -o skill.zip -d .claude/skills/catvton-train-reconstruction-hello-hoy && rm skill.zipInstalls to .claude/skills/catvton-train-reconstruction-hello-hoy
Activation
This is the description your AI agent reads to decide when to run this skill — the better it matches your request, the more reliably it fires.
Use when implementing, reviewing, or extending CatVTON-style training in this repository. Focus on DressCode-based training, agnostic mask generation from local dataset annotations, and attention checkpoint export compatible with CatVTON inference layouts.About this skill
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
- Read
references/repo-notes.md. - Keep the implementation scoped to this repo before reaching back into the original CatVTON repo.
- Preserve CatVTON-compatible checkpoint layout:
<output>/<dataset_tag>/attention
- Prefer self-attention-only training unless the user explicitly asks for a different fine-tuning target.
- When changing dataset logic, validate against the actual files under
data/DressCode.
Validation rules
- Confirm
train_pairs.txtand 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