Video Generative Priors for 3D
Definition
The use of large video diffusion models (pretrained on natural temporal motion) as structural priors to regularize, denoise, and restore consistency to 3D neural scene reconstructions.
Key Characteristics
- Enforces temporal and multi-view consistency across camera trajectories.
- Eliminates floating artifacts and view-dependent blurry hallucinations in NeRFs and 3D Gaussian splatting.
- Preserves underlying 3D geometry while enhancing surface texture resolution.
Applications
- 3D Gaussian Splatting refinement
- Photogrammetry restoration
- Virtual production
Mentions in Source
- βUses video diffusion temporal attention as a prior to fix floaters, blurry textures, and geometric inconsistencies in 3D scene captures.β β fixanything-3d-consistent-rendering-refinement