See-through

See-through is an open-source research framework for turning a single static anime character illustration into editable, fully inpainted RGBA body-part layers with an inferred drawing order. It was introduced in a 2026 paper by Jian Lin, Chengze Li, Haoyun Qin, Kwun Wang Chan, Yanghua Jin, Hanyuan Liu, Stephen Chun Wang Choy, and Xueting Liu. The authors list affiliations with Saint Francis University, the University of Pennsylvania, Spellbrush, and Shitagaki Lab; the official source repository is published through the Shitagaki Lab GitHub organization.¹²

The framework addresses a preparation step in 2.5D character animation. Artists ordinarily separate a flat illustration into parts, paint content that was hidden in the original view, and specify which fragments should appear in front of others. See-through attempts to automate that work while preserving the source illustration’s line work, color, and overall appearance.¹

Output and scope

See-through generates transparent layers for semantic regions such as hair, face, eyes, clothing, limbs, and accessories. The paper describes a 19-class body-part taxonomy, while the repository’s V3 inference pipeline can stratify its results into as many as 23 layers. Its main script exports a layered Photoshop document together with intermediate segmentation masks and depth maps.¹²

The result is intended to be a 2.5D-ready representation, not a finished Live2D model. See-through does not create deformation meshes, physics parameters, or motion curves, and it does not decide how a character should move. Those rigging and animation stages still require downstream software and artistic work.²

Method

Training data

The researchers constructed supervision from existing Live2D models. A Live2D character is composed from textured ArtMeshes whose visibility masks and artist-defined drawing-order indices provide precise fragment boundaries and occlusion relationships. Because mesh names and hierarchies are not standardized across creators, the data engine assigns the fragments to a fixed semantic taxonomy.¹

The labeling process begins with image tags and Grad-CAM activation maps, snaps the coarse responses to ArtMesh visibility masks, and refines the result with class-specific Segment Anything decoders. Predictions are projected back onto the meshes, propagated to fully hidden fragments, and manually verified. The recorded drawing-order indices are normalized into pseudo-depth values that represent compositing order rather than physical distance from a camera.¹

Layer generation

The layer generator is based on Stable Diffusion XL. A transparency decoder extends the model’s RGB latent representation into RGBA output. Training proceeds in two stages: the first teaches the model to extract one requested body part, and the second denoises the complete stack jointly. A Body Part Consistency Module lets the predicted parts exchange information so that ambiguous content is allocated more consistently across the whole decomposition.¹

See-through uses a separately fine-tuned Marigold diffusion model to predict pseudo-depth for the layers. Selected categories can then be divided into front and back strata, allowing a single semantic part such as hair to pass both behind and in front of the face. An inpainting step fills regions exposed by that subdivision before the layers are assembled into the final stack.¹

Distribution and operation

The upstream repository provides inference and model-training code under the Apache License 2.0, with model checkpoints distributed separately through Hugging Face. Its primary inference script accepts either one image or a directory and writes the layered PSD and intermediate artifacts to a workspace directory.²³

The normal pipeline uses BF16 precision and, at the documented 1280-pixel working resolution, requires approximately 12–16 GB of GPU memory. The repository also includes group-offload, NF4 quantization, and block-swap paths for lower-memory GPUs. These modes trade speed, precision, or system-memory use for lower peak video-memory demand.²

Limitations

Content hidden in the input image cannot be observed directly. See-through therefore synthesizes a plausible completion rather than recovering ground-truth artwork. Fine hair tips, small decorations, accessories, and other high-frequency details may remain ambiguous or show model-generated visual patterns. In the paper’s informal artist review, participants generally treated the output as a useful starting point but still identified cases requiring comparison with the original and manual editing.¹

Pseudo-depth likewise records a useful relative layer order, not metric 3D geometry. Even a visually faithful reconstruction can need different separations for a particular animation, such as distinct left and right limbs. The output should therefore be reviewed and adapted to the intended rig rather than treated as an automatic replacement for a Live2D artist.¹²

Relevance to return moe

return moe maintains an independent fork of See-through to build Docker images and deployment tooling optimized for cloud GPU providers such as RunPod. The fork adds a web interface, container packaging, and cloud-oriented operational guidance.⁴

References

Footnotes

1. Jian Lin et al., See-through: Single-image Layer Decomposition for Anime Characters, arXiv:2602.03749 (2026). ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹

2. See-through source repository, Shitagaki Lab. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

3. See-through source-code license, Apache License 2.0. ↩

4. return moe’s See-through fork and its RunPod deployment guide. ↩