See-Through turns anime art into layered PSDs

Anyone who has ever rigged a Live2D model or animated a flat illustration knows the worst part of the job: layer separation. You get a flattened piece of art, and you have to spend hours meticulously cutting out bangs, eyes, and sleeves, then manually painting back whatever was hidden underneath.

An open-source project called See-Through just automated that headache. Based on an accepted SIGGRAPH 2026 paper, it transforms static anime art into manipulatable 2.5D assets in minutes.

From a flat JPEG to 23 layers

Feed See-Through a single anime illustration, and it slices the image into up to 23 distinct semantic layers, exporting a fully structured PSD file with inferred drawing order.

The layers cover everything from hair, facial features, and eyes to clothing and accessories. What makes it genuinely useful is its inpainting capability. If a character's arm blocks their torso or furniture, See-Through generates the hidden surfaces so you do not end up with awkward hollow cutouts.

Under the hood, the pipeline uses LayerDiff 3D to handle transparent layer generation alongside a fine-tuned Marigold model for pseudo-depth estimation. Beyond the PSD, it exports depth maps and masks, and includes helper scripts to further slice elements by depth or left-right symmetry.

Running it on budget GPUs

Hardware requirements are refreshingly reasonable. The default bf16 pipeline runs at 1280 resolution and needs around 12 to 16 GB of VRAM.

For 12 GB cards, enabling group offload cuts peak memory to roughly 10 GB at the cost of about 1.5x slower processing. If you are on an 8 GB card, the repository includes an NF4 4-bit quantized pipeline that squeezes right into 8 GB with virtually no noticeable drop in visual quality. You can also dial the resolution down to 1024 for extra breathing room.

If you prefer not to touch terminal commands, the team set up a Hugging Face Space with ZeroGPU allowing a couple of runs per day. Users in mainland China can also use a free ModelScope web demo that supports slightly higher resolutions.

Why it matters

For indie animators, VTuber riggers, and 2D game developers, See-Through tackles the most tedious bottleneck in the pipeline. Instead of spending an entire afternoon tediously slicing and redrawing occluded limbs, artists get a reasonably clean 23-layer PSD to start with, turning hours of prep work into a quick cleanup pass.

We like this approach because it does not try to reinvent the wheel with brittle end-to-end video generation. By focusing on standard PSD output and depth data, it slots directly into traditional workflows across Photoshop, After Effects, and Live2D Cubism without forcing creators to change tools.

That said, 23 layers will not immediately produce a broadcast-ready VTuber avatar, which typically requires dozens of micro-slices just for facial tracking. But for dynamic game dialogue portraits, parallax wallpapers, and 2.5D motion graphics, it gets you 80 percent of the way there almost instantly.