jtydhr88/ComfyUI-See-through

What it solves

This project provides a way to turn a single anime illustration into a layered 2.5D model. It solves the problem of manually cutting out character parts (like hair, eyes, and clothing) to create assets for Live2D or 3D parallax animations, which is traditionally a labor-intensive process.

How it works

The plugin uses a combination of two AI models: LayerDiff (an SDXL-based pipeline) to generate transparent semantic layers and Marigold (a fine-tuned depth estimation model) to determine the correct drawing order of those layers. It can split symmetric parts like eyes and ears into left and right versions and uses depth clustering to separate front and back hair. The final output is exported as a layered PSD file directly through the browser.

Who it’s for

Digital artists, anime illustrators, and Live2D riggers who need to automatically decompose images into layers for animation workflows.

Highlights

• Automatic Layer Decomposition: Generates up to 24 semantic transparent layers from one image.
• Depth-Aware Ordering: Uses fine-tuned depth maps to ensure layers are stacked correctly.
• Symmetric Splitting: Automatically separates left/right eyes, ears, and handwear.
• PSD Export: Direct browser-based download of layered PSD files.
• VRAM Optimization: Includes features like tag embedding caching and group offloading to support lower-end GPUs.