shitagaki-lab/see-through

A research pipeline worth trying if you already have anime artwork and need separated layers; it is not a one-click consumer tool, and the VRAM and setup requirements are real.

The problem

A single flat illustration of an anime character is usually one canvas: hair, face, eyes and clothing are fused together. If you want to move a fringe, swap a mouth or rig the character for animation, you first have to repaint every hidden region and decide which layer sits above which. That masking and inpainting work is slow, repetitive and easy to get wrong at the edges. See-through changes that step: you give it one image, and it generates a layered PSD with up to 23 semantic layers, each filled in and ordered by inferred drawing order project README. The result is an editable source file rather than a stack of half-transparent cutouts.

What's different

The mechanism is a two-model pipeline. A diffusion model, LayerDiff 3D, generates transparent layers, while a Marigold depth model fine-tuned for anime supplies pseudo-depth model list. The pipeline uses that depth and body-part grouping to split head and body into semantic pieces such as hair, eyes, clothing and accessories, then writes a layered PSD with drawing order. The practical benefit is that the difficult step, reconstructing occluded areas, happens inside the decomposition instead of in a separate manual inpainting pass. The README notes that head and body separation run as two continuous stages, which can take longer than the original model in the paper acknowledged. The main trade-off is that this is a research release, not an editing application: layer quality, and how much retouching a layer needs, are things you must judge on your own artwork rather than assume.

The project has attracted third-party integrations: a ComfyUI node with in-browser PSD export, a desktop tool that generates blink and lip-sync materials, and StretchyStudio, a browser puppet-animation editor that accepts the PSD output and auto-rigs it community integrations. Those show the intended downstream workflow: decompose, then animate. They do not establish that the initial decomposition is artifact-free.

Who should try it

A good fit is a 2D artist or Live2D-style rigger with an existing library of anime illustrations who wants to try automated layer separation before committing hours to manual masking, and who has a GPU in the 8–16 GB range. A small studio exploring a semi-automated pipeline, or someone already using ComfyUI or StretchyStudio and wanting to feed PSD layers into it, is also a reasonable candidate. A poor fit is someone without a GPU who expects a polished desktop app: there is no local GUI described, only scripts and notebooks. For a quick evaluation without setup, the hosted Hugging Face Space allows roughly 1–2 PSD extractions per day at 1280 resolution after registering, and the ModelScope demo is presented as free with slightly higher resolution online demos. Those quotas make the hosted route unsuitable for batch work.

Requirements and limitations

The default pipeline runs at bf16 and needs roughly 12–16 GB of VRAM at 1280 resolution. A 12 GB card can use group offload to bring peak usage to about 10 GB, and an 8 GB card can use an NF4 quantized pipeline at about 8 GB, with lower resolution as a further lever VRAM guidance. Setup involves conda, pinned PyTorch CUDA 12.8 wheels, with notes for aarch64 and AMD ROCm users, plus downloading the model weights from Hugging Face environment setup. Ongoing upkeep means tracking model downloads and PyTorch/ROCm compatibility rather than a packaged updater.

The harder limitation is qualitative: the review found no supplied benchmark of layer accuracy across styles or resolutions, and the two-stage head/body split may take longer than the paper's original model. You should also expect to inspect each exported PSD and refine edges or hidden regions. This is a conference research project; the README explicitly warns that paid sites offering this functionality are not affiliated with it project README.

HotRepo's take

On 2026-10-28 the project had 5,062 stars, with +78 recorded that day on the daily trending board, its only tracked board appearance in the supplied sample repository. That is a single dated snapshot from a repository created on 2026-03-31, so trend persistence cannot be judged from it. The sensible decision is to evaluate it against your own images: start with the hosted demo, then check whether your card fits the documented 8–16 GB ladder before installing the full pipeline. Verify one specific compatibility point first: whether the pinned PyTorch wheel matches your platform, since aarch64 and ROCm users are directed to substitute versions.