VGGT-Diff: Visual Geometry Meets Diffusion for Sparse-View Novel View Synthesis
The paper routes explicit 3D geometry into a pretrained video diffusion model to improve sparse-view novel view synthesis.
VGGT-Diff attaches each visual token to a 3D point and confidence score, then uses a Visual Geometry Router to turn that evidence into query-aligned latent conditions. The model also regularizes predicted residuals along reliable 3D tracks to improve consistency across target views. The authors report competitive or state-of-the-art results on interpolation and extrapolation across different viewpoint difficulties. HF Daily Papers' note
VGGT-Diff attaches each visual token to a 3D point and confidence score, then uses a Visual Geometry Router to turn that evidence into query-aligned latent conditions. The model also regularizes predicted residuals along reliable 3D tracks to improve consistency across target views. The authors report competitive or state-of-the-art results on interpolation and extrapolation across different viewpoint difficulties. HF Daily Papers' note
score 5