RFD 1039: Model image for trellis2_image_mesh_painting
Problem
Mesh painting takes a mesh that exists and gives it a texture from an image. It uses the same TRELLIS.2 weights as RFD 1038. A second copy of those weights costs 8.0 GB and buys nothing.
Decision
Build FROM the RFD 1038 image. Add a predict.py, and add no weights. The model image is then about 40 MB, and not 8 GB.
See DETAILS.md for the model’s shared-weight cost, the predict() interface, and why the UV unwrap must not hide inside this model image.