README.md

metadata

title: 'See-through: Layer Decomposition'
emoji: 🔍
colorFrom: blue
colorTo: purple
sdk: gradio
app_file: app.py
pinned: false
license: mit

See-through: Single-image Layer Decomposition for Anime Characters

GitHub | Paper (arXiv:2602.03749)

Upload an anime character illustration to decompose it into fully-inpainted semantic layers with depth ordering, exported as a layered PSD file.

How it works

This demo runs the full See-through pipeline:

1. LayerDiff 3D — diffusion-based transparent layer generation (SDXL)
2. Marigold Depth — pseudo-depth estimation fine-tuned for anime
3. Post-processing — layer stratification and PSD export

Usage

1. Upload a character illustration (square images work best; non-square images are center-cropped)
2. Optionally adjust the seed
3. Click Submit and wait (~1-2 minutes)
4. Download the layered PSD file

Notes

• This demo is for PSD inference only (layer decomposition). For the full toolkit (body parsing, heuristic stratification, UI, training), see the GitHub repo.
• Runs on ZeroGPU — free for all users, but GPU quota is limited. PRO users get higher quota and priority. If you experience timeouts during peak hours, please try again later or run locally.
• Disclaimer: This demo uses a newer model checkpoint and may not fully reproduce identical results reported in the paper.