RFD 1040: Model image for pixal3d_image_to_textured_mesh
Problem
Pixal3D is the image to 3D path in daily use. It writes a metal map and a roughness map, which TRELLIS.2 does not.
This RFD first recorded its parameter count and its license as unknown, and it asked for a measurement on the DGX. There is no DGX. The measurement came from the published checkpoints instead.
Decision
Package Pixal3D as the primary image to 3D worker. Upstream is TencentARC/Pixal3D, and it uses the MIT license.
It is a plain Docker image that serves HTTP, and not a Cog. RFD 1036 records why: vast.ai rents an instance and runs a container on it.
Call the upstream inference.py, and do not reimplement the cascade. A copy here would drift from the commit this image pins.
See DETAILS.md for the Docker contract-stage test, the checkpoint measurement, and the format choice. It also gives the predict() interface, the build-time downloads, and what this RFD corrects.