shahafvl/gaussianformer-256px

GaussianFormer, 256 px training checkpoint

The checkpoint after the two 256 px training stages (Gaussian input head, then the whole model, both from RenderFormer's weights). It is the starting point of the last two stages that produce GaussianFormer, and is published so those stages can be reproduced exactly. For rendering, use shahafvl/gaussianformer.

Code: github.com/SVLwoof/gaussianformer

BASE=shahafvl/gaussianformer-256px scripts/train.sh   # stages 3 (256 px recovery) and 4 (512 px)

Architecture: RenderFormer's layout with 8 px patches and dense cross-attention; stage 3 switches on the projected 2-D RoPE, the windowed cross-attention and 4 px patches.

License

Trained on Objaverse_Splats, a subset of Objaverse whose terms restrict commercial use; CC-BY-NC-4.0, research and non-commercial use only.