See-through: Single-image Layer Decomposition for Anime Characters

TL;DR

See-through automates the conversion of a single anime illustration into a 2.5D-manipulable, layered PSD with up to 23 fully-inpainted semantic layers (hair, face, eyes, clothing, accessories, etc.) and inferred drawing order [§Abstract]. Two coupled diffusion-based components do the work: a Body Part Consistency Module (LayerDiff 3D, SDXL-based) generates transparent RGBA layers under global geometric constraints, and an anime-fine-tuned Marigold pseudo-depth model resolves stratification of intricate structures like interleaving hair strands [§Abstract; README]. Supervision comes from a scalable data engine that bootstraps pixel-perfect layer ground truth (semantics + occluded geometry) from commercial Live2D model files, sidestepping the per-image manual segmentation that currently dominates the workflow [§Abstract]. Conditionally accepted to SIGGRAPH 2026; full code, models, and a HuggingFace ZeroGPU demo are open under research-license terms.

Key claims

Single-image decomposition into up to 23 semantically distinct, fully-inpainted layers with inferred drawing order — eliminating the manual segmentation + occluded-region hallucination steps in the professional Live2D workflow [§Abstract; README].
A diffusion-based Body Part Consistency Module enforces global geometric coherence across the per-layer transparent generations, preventing per-part drift that pure layerwise inpainting suffers from [§Abstract].
A pixel-level pseudo-depth inference mechanism (Marigold fine-tuned for anime) resolves stratification of overlapping fine structures, explicitly called out for “interleaving hair strands” [§Abstract].
Training supervision is bootstrapped from commercial Live2D models rather than manual annotation: pixel-perfect semantics and the artist’s own hidden-geometry choices serve as ground truth, addressing data scarcity for this task [§Abstract; README acknowledgements].
Practical deployment envelope: bf16 inference at 1280 resolution runs in ~12–16 GB VRAM; NF4-quantized variant runs in ~8 GB with PSNR ~30 dB / SSIM ~0.96 vs the bf16 baseline [README “Low-VRAM Users”].
The authors explicitly disclaim this is a complete Image-to-Live2D pipeline: rigging (deformation mesh, physics, motion curves) and Live2D-artistic decomposition choices remain manual; See-through is positioned as the “tedious-step automation” upstream of rigging [README “Discussion: Is this Image-to-Live2D?”].

Method

A single anime illustration enters a two-stage pipeline. Stage one is LayerDiff 3D, an SDXL-based diffusion model that generates transparent RGBA layers conditioned on body-part tags from a separate SAM-based body parsing model; the diffusion is regularized by a Body Part Consistency Module enforcing global geometric coherence across layers [§Abstract; README Models]. Stage two is an anime-fine-tuned Marigold monocular depth model that produces pseudo-depth at pixel resolution, which the pipeline uses to stratify layers (especially fine, interleaving structures) and to optionally apply downstream depth-based / left-right splits [§Abstract; README Inference Scripts, heuristic_partseg.py]. The output is a layered PSD with up to 23 named body-part layers in the V3 model, plus intermediate depth maps and segmentation masks [README]. Training data is generated by a scalable engine that parses commercial Live2D model files (via the companion CubismPartExtr repo) to extract per-part textures, drawing order, and hidden-geometry inpaintings as ground truth [§Abstract; README Training; preprint mentions 23 pages / 20 figures].

Results

The paper is conditionally accepted at SIGGRAPH 2026 Conference Proceedings [README]. Headline numbers were not extractable from the abstract / repo at filing time (the arxiv HTML page failed to fetch); the only quantitative figures in the repo are deployment-side: NF4 quantized inference matches bf16 at PSNR ~30 dB, SSIM ~0.96 at 1280 resolution [README “Low-VRAM Users”]. The tweet announcing the open-source release reports 536K views and ~4.5K engagement at the time of fetch, indicating notable community uptake; a ComfyUI integration (@jtydhr88) and a downstream rigging tool (StretchyStudio) have already been built on top of the PSD output [README “Community Support”].

Why it’s interesting

This is the first anime/illustration-specific entry in the wiki’s layered-decomposition cluster (Layered Image/Video Decomposition). The existing papers there target either natural photographs (Generative Image Layer Decomposition with Visual Effects (LayerDecomp) — LayerDecomp’s BG + RGBA-FG-with-effects formulation, which explicitly struggles on “photographic complex lighting/shadows” per Qwen-Image-Layered-Control: text-guided layer extraction) or graphic-design templates (ART: Anonymous Region Transformer for Variable Multi-Layer Transparent Image Generation — ART with 50+ layers, anonymous regions, layout-conditional 3D RoPE; LICA: A Dataset of 1.5M Layered Graphic Design Compositions — LICA template-native dataset). See-through occupies a third regime — line-art / cel-shaded illustrations with deterministic occluded geometry — and bypasses the photo-supervision problem by harvesting ground truth from commercial Live2D model files instead of relying on counterfactual photography (LayerDecomp’s hybrid) or template scraping (LICA/ART’s MLTD). The 23-layer depth-stratified output also lines up with the “joint-generator route to layered output” question raised in Layered Image/Video Decomposition‘s open questions: this paper instead solves it by training a decomposer with very high-quality artist-supplied ground truth — orthogonal to the ART: Anonymous Region Transformer for Variable Multi-Layer Transparent Image Generation route of training a joint generator on a curated multi-layer dataset. For Luma, the 2.5D-from-single-image angle is also adjacent to the broader Image-to-3D / NVS direction without committing to full 3D — useful reference point for anime/illustration verticals.