RFD 1006: Layer decomposition (See-Through)
Problem
An anime illustration is one flat image. A user wants to edit the hair, the face, and the clothing separately. Image to 3D also works better with a clean subject.
Decision
Add a layer decomposition stage between text to image and image to 3D. The stage runs the See-Through model.
The task decomposes one image into RGBA body-part layers. It returns a layers zip, a PSD, a composite URL, and a layer count. The Studio pipeline stores the artifacts on the layer_decomposition node. Image to 3D prefers the composite image as its input.
References
- Paper: https://doi.org/10.1145/3799902.3811209
- Cog model: weftspun/see-through (branch cog)
- Executor:
src/library/taskManager.js - URLs:
src/library/taskModelUrl.js
WITHDRAWN 2026-08-29: See-Through is a reference, not a model here. RFD 1166 dropped it from the candidate ranking because every checkpoint its inference scripts load states no licence, and the depth one is a fine-tune of an OpenRAIL++-M model whose restrictions travel into derivatives. What is kept is the layer taxonomy in common/live2d/scrap_model.py. The approach below still describes the task; it no longer describes a model this workspace can run.