shitagaki-lab/see-through

What it solves

See-through automates the process of turning a static 2D anime illustration into a manipulatable 2.5D model. It solves the tedious manual work of segmenting a character into different parts and inpainting the hidden areas behind those parts, which is typically required for animation or rigging.

How it works

The framework decomposes a single image into up to 23 semantically distinct, fully inpainted layers (such as hair, eyes, and clothing) with inferred drawing orders. It achieves this using a pipeline that combines:

• LayerDiff 3D: A diffusion-based model (SDXL) for generating transparent layers.
• Marigold Depth: A fine-tuned depth estimation model specifically for anime to determine layer order.
• SAM Body Parsing: A semantic segmentation model to identify specific body parts.

The final output is a layered PSD file that can be imported into animation or rigging tools.

Who it’s for

This tool is designed for anime artists, 2.5D animators, and VTuber creators who want to accelerate the preparation of character assets for rigging and animation.

Highlights

• High Granularity: Decomposes characters into up to 23 distinct semantic layers.
• Automated Inpainting: Fills in occluded regions so layers are complete even when moved.
• PSD Export: Outputs standard layered files compatible with professional design software.
• VRAM Optimizations: Includes NF4 quantization and group offloading to support GPUs with as little as 8GB VRAM.
• Research-Backed: Published in ACM SIGGRAPH 2026.