Upload a single anime character illustration to automatically decompose it into fully-inpainted semantic layers with depth ordering, exported as a layered PSD file.
A one-click portable launcher based on See-through (Apache 2.0 License).
- Automatically decompose anime character images into up to 23 semantic layers:
- Hair: front hair, back hair
- Head: head, face, nose, mouth
- Eyes: eyewhite, irides, eyelash, eyebrow
- Accessories: headwear, eyewear, earwear, neckwear
- Body: ears, neck
- Clothing: topwear, bottomwear, legwear, footwear, handwear
- Other: tail, wings, objects
- Every layer is fully inpainted, not simply cropped
- Automatic depth ordering for each layer
- Export as PSD file
- Gradio web interface with English / Chinese toggle
- OS: Windows 10 / 11
- Python: 3.10 - 3.12 (if not installed,
run.batwill automatically install Python 3.12 via winget) - GPU: NVIDIA GPU with at least 8 GB VRAM (GTX 10 series to RTX 50 series supported)
- NVIDIA Driver: Version 560 or above (latest version recommended)
- Disk Space: ~20 GB (including model downloads)
- Download zip from Releases and extract, or clone this repository
- Double-click
run.bat - First run will automatically create a virtual environment and install all dependencies (~10-20 minutes)
- Browser will automatically open the Gradio interface
- Upload an image and click "Start Processing"
Warning
The first time you process an image, models will be downloaded automatically (~13 GB). They will not be re-downloaded afterward.
If automatic download is too slow or your network is unstable, you can download models manually.
This project uses two HuggingFace models:
| Model | Size | Link |
|---|---|---|
| LayerDiff (Layer Decomposition) | ~9.5 GB | layerdifforg/seethroughv0.0.2_layerdiff3d |
| Marigold (Depth Estimation) | ~3.3 GB | 24yearsold/seethroughv0.0.1_marigold |
Run run.bat once to set up the virtual environment, then open a command prompt:
cd your\path\see-through-portable
venv\Scripts\activate
huggingface-cli download layerdifforg/seethroughv0.0.2_layerdiff3d --cache-dir models/hub
huggingface-cli download 24yearsold/seethroughv0.0.1_marigold --cache-dir models/hub| Parameter | Default | Description |
|---|---|---|
| Random Seed | 42 | Different seeds produce different decomposition results |
| Resolution | 1280 | Higher = better quality but slower and more VRAM. Image is center-padded to square |
| Inference Steps | 30 | Denoising steps. More = better quality but slower. Not recommended to change |
| Left/Right Split | OFF | Split gloves, eyes, ears, etc. into separate left/right layers |
| Cache Tag Embeddings | ON | Pre-compute text embeddings and unload text encoders, saves ~2 GB VRAM with zero speed penalty |
| Group Offload | OFF | Move model blocks on/off GPU as needed. Drastically reduces VRAM but 2-3x slower |
| Depth Resolution | -1 | -1 = same as layer resolution. Lower values save VRAM with slightly reduced depth accuracy. Custom default: 768 (model training resolution) |
12 GB+ VRAM (e.g. RTX 3060 12G, RTX 4070 and above): Default settings are fine. Cache Tag Embeddings is already enabled by default.
8-12 GB VRAM (e.g. RTX 3060 8G, RTX 4060): Try the following in order, from least to most impact on speed:
- Cache Tag Embeddings = ON (already enabled by default) — Saves ~2 GB with zero speed penalty
- Lower Depth Resolution — Uncheck "Depth resolution same as layers", defaults to 768 (model training resolution), adjustable. Saves VRAM with slightly reduced depth accuracy
- Lower Resolution — e.g. 1024 instead of 1280, reduces both VRAM and computation time
- Group Offload = ON — Last resort. Drastically reduces VRAM but 2-3x slower
Output files are located in the workspace/layerdiff_output/ folder:
<image_name>.psd— Multi-layer PSD file<image_name>/— Individual layer PNG files
Q: run.bat closes immediately?
A: Right-click run.bat > Edit, check that the file encoding is UTF-8 with BOM or ANSI. Or run run.bat directly in cmd to see error messages.
Q: run.bat is blocked by Windows SmartScreen? A: Click "More info" → "Run anyway".
Q: "No compatible Python (3.10 - 3.12) found"?
A: run.bat will automatically install Python 3.12 via winget. After installation, close the window and double-click run.bat again.
Q: How to uninstall the auto-installed Python?
A: Open a command prompt and run: winget uninstall Python.Python.3.12
Q: "No NVIDIA GPU with CUDA detected"? A: Make sure you have NVIDIA driver version 560 or above. Download the latest driver: https://www.nvidia.com/download/index.aspx . AMD GPUs are not supported.
Q: "CUDA error: no kernel image is available"? A: Your GPU is too old. This tool requires at least an NVIDIA GTX 10 series GPU.
Q: C++ compiler error during dependency installation? A: Install Visual Studio Build Tools.
Q: Model download was interrupted?
A: Run run.bat again. The download will resume from where it left off.
Q: Does the folder path with spaces or special characters cause issues?
A: It is recommended to extract the package to a simple path, e.g. D:\see-through-portable, avoiding paths with spaces or special characters.
Q: How long does it take to process one image? A: Processing time varies greatly depending on GPU performance and image resolution.
This project is based on See-through by shitagaki-lab, licensed under Apache 2.0.