shahafvl/gaussianformer
Image-to-Image • 0.2B • Updated • 100 • 1
Transformer rendering of 3D Gaussian Splats, adapted from RenderFormer. gaussianformer = current flagship; versioned checkpoints below.
Note Current flagship (V22)
Note V22: dense cross-attention with the projected footprint (same weights as the flagship)
Note V21: dense cross-attention, no footprint (for comparison)
Note V20: windowed cross-attention (the earlier flagship)
Note 256 px training checkpoint (start of training stages 3-4)
Note V17b
Note V10b