3D-GIMP slashes inpainting cost with single-view diffusion and PatchMatch
No more per-view diffusion: 3D-GIMP renders 3D scenes faster with better consistency.
Researchers Tian, Schmalstieg, and Mori present 3D-GIMP, a hybrid paradigm for high-fidelity object removal in 3D Gaussian Splatting. Rather than diffusing every view, it performs a single generative inpainting on a key reference view as an appearance prior, then uses a 3D-aware PatchMatch algorithm to propagate textures to all remaining views via correspondence matching. This bypasses the stochastic nature of frame-by-frame diffusion, addressing hallucination drift and multi-view inconsistencies. 3D-GIMP maintains high-frequency details across arbitrary resolutions while ensuring mathematically consistent 3D reconstruction. Experiments show it achieves competitive inpainting quality compared to prior multi-view diffusion methods and outperforms them in rendering speed and view consistency.
- Uses only one diffusion inpainting pass on a key reference view instead of per-view diffusion
- 3D-aware PatchMatch algorithm propagates textures to all other views, eliminating hallucination drift
- Matches quality of multi-view diffusion methods while achieving faster rendering and better view consistency
Why It Matters
3D-GIMP makes 3D scene editing practical for professionals by cutting compute while preserving high-frequency detail and multi-view consistency.