See-through is a Python framework that decomposes single anime character images into layered 2.5D models with inferred drawing order.
The tool addresses the challenge of converting static anime illustrations into manipulatable multi-layer representations. It works by analyzing a single image and automatically decomposing it into fully inpainted, semantically distinct layers representing different character elements such as hair, face, eyes, clothing, and accessories. The framework can produce up to 23 layers per image and infers the drawing order between them. The approach combines diffusion-based layer generation through a model called LayerDiff 3D, pseudo-depth estimation fine-tuned for anime using Marigold Depth, and semantic segmentation capabilities including SAM-based body parsing and optional instance segmentation.
This tool suits developers and artists working with anime content who need to programmatically extract and manipulate individual character components from flat images. It is particularly valuable for animation pipelines, character rigging workflows, and interactive media applications where layer separation and depth information are prerequisites. The project provides optional annotator tiers for different segmentation capabilities, allowing users to install only the dependencies they need for their specific use case.
The maintainers respond to new issues and pull requests within a day, indicating active engagement with the codebase and user feedback.
