RFD 1038 details: the model, the interface, staging

The model

Property	Value
Parameters	4.0 B, estimated
bf16	8.0 GB
Q4_K_M	2.20 GB
License	MIT
Format	bf16, per RFD 1027

The interface

predict() takes the image, the texture resolution, and the face budget. It returns the base USD layer, and a GLB beside it. RFD 1053 gives that rule.

Input	Type	Default
image	Path	none
texture_resolution	int	1024
decimation_target	int	210000
seed	int	-1

decimation_target must not exceed 210000. That is API_MAX_MESH_VERTICES in src/library/aiModelsCatalog.js, and it matches the API upload cap. A larger mesh fails the next stage.

Two stages, one container

The sparse structure flow runs first, and the SLat flow runs second. Both stay in one model image. They share the DINOv2 image encoder, thus a split would load that encoder twice.