shitagaki-lab /see-through

"Single-image Layer Decomposition for Anime Characters" (SIGGRAPH 2026 Conference Paper)

★ 5,681⑂ 489👁 41⑃ 16● 35▤ 0
PythonApache-2.0Created 2026-03-31
+12
Daily
5,681
stars
489
forks
41
watchers
16
prs
35
issues
0
releases

README

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







_Jian Lin1, Chengze Li1, Haoyun Qin2,3,4, Kwun Wang Chan1, Yanghua Jin3, Hanyuan Liu1, Stephen Chun Wang Choy1, Xueting Liu1_

1Saint Francis University   2University of Pennsylvania   3Spellbrush   4Shitagaki Lab

Corresponding author

Published in ACM SIGGRAPH 2026 Conference Papers.



---

 Notice: This is an open-source research project. We have not set up any paid service for this tool. If you encounter a website charging for this functionality, it is not from us. Use at your own risk.

 声明: 本项目为开源研究项目,我们未开设任何付费服务。如遇到以此功能收费的网站,均与我们无关,请注意甄别。

 TL;DR

We introduce a framework that automates the transformation of static anime illustrations into manipulatable 2.5D models. Our approach decomposes a single image into fully inpainted, semantically distinct layers with inferred drawing orders — up to 23 layers including hair, face, eyes, clothing, accessories, and more.

!Our Representative Image



  

https://github.com/user-attachments/assets/023d271f-d8d7-4f6b-9083-96e714fb93e0


  
  This is our trailer video. Click to play.


 Environment Setup

bash
 1. Create environment
conda create -n see_through python=3.12 -y
conda activate see_through

 2. Install PyTorch (CUDA 12.8)
 aarch64 users: the pinned versions below may not be available; use torch=2.9.0 instead
pip install torch==2.8.0+cu128 torchvision==0.23.0+cu128 torchaudio==2.8.0+cu128 \
  --index-url https://download.pytorch.org/whl/cu128

 AMD ROCm users: install the ROCm wheel set that matches your local ROCm runtime instead.
 For example, on ROCm 7.2:
 pip install torch torchvision torchaudio \
   --index-url https://download.pytorch.org/whl/rocm7.2

 3. Install dependencies (includes common utilities and annotators)
pip install -r requirements.txt

 4. Create assets symlink (you can also copy assets to the root if you prefer)
ln -sf common/assets assets


Optional annotator tiers (install as needed):

| Tier | Command | What it adds |
|------|---------|-------------|
| Body parsing | pip install --no-build-isolation -r requirements-inference-annotators.txt | detectron2 for body attribute tagging |
| SAM2 | pip install --no-build-isolation -r requirements-inference-sam2.txt | SAM2 for language-guided segmentation |
| Instance seg | pip install -r requirements-inference-mmdet.txt | mmcv/mmdet for anime instance segmentation |

 Note: Always run scripts from the repository root as the working directory.

 Scripts & Models

 Models

| Model | HuggingFace Repo | Description |
|-------|-----------------|-------------|
| LayerDiff 3D |  | Diffusion-based transparent layer generation (SDXL) |
| Marigold Depth |  | Pseudo-depth estimation fine-tuned for anime |
| SAM Body Parsing |  | Semantic body part segmentation|

 Inference Scripts

| Script | Purpose |
|--------|---------|
| inference/scripts/inference_psd.py | Main pipeline — end-to-end layer decomposition → PSD output |
| inference/scripts/syn

Top issues

Snapshot: 2026-11-16T09:23:14Z · Data license Apache-2.0 · JSON