See-through

See-through is a layer decomposition workflow for anime character art rather than a general image editor. You give it one illustration, and it tries to split the character into fully inpainted semantic layers with depth order, then exports the result as a PSD. The useful angle is workflow clarity: there is a public Space for a fast proof, while the public repo and checkpoints document the same PSD inference path locally. A good first test is one centered character with visible occlusion around hair, sleeves, or accessories, because that reveals quickly whether the layer stack is genuinely useful or still too messy to trust.

Start with the public Space and use one clean anime character image instead of a busy scene or extreme crop. On the first run, look for three things: whether hidden regions are filled plausibly, whether the layer order makes editing sense, and whether smaller details like hair strands or accessories survive the split instead of collapsing together. If it passes that check, move to the public repo and run `inference/scripts/inference_psd.py` locally so you can control resolution, VRAM tradeoffs, and the exported PSD output more reliably.

Treat this as PSD decomposition, not a finished animation pipeline. It is specific to anime character art, the local path is still GPU-heavy, and the output can still need manual cleanup before serious rigging or production use.

• Break one anime character illustration into editable PSD layers before manual cleanup.
• Prototype Live2D or layered animation prep without segmenting every part by hand from scratch.
• Test whether occluded regions are reconstructed well enough for character edits or motion experiments.
• Compare an open layer-decomposition workflow against manual Photoshop masking on your own art.

Try the demo

View model page