FineSplat uses local grouping for detailed 3D Gaussian splats
Original title: FineSplat: Fine Patches, Local Groups, Detailed Splats
View-wise and 3D grouping plus rendering-error feedback keep Gaussian counts at or below prior methods.
FineSplat is a feed-forward 3D Gaussian Splatting framework built on fine image patches. View-wise grouping and 3D spatial grouping avoid costly global multi-view attention while reconnecting geometrically related tokens. Reconstruction error supplies an explicit rendering signal, and a compact Gaussian scheme limits primitive count. The authors report state-of-the-art novel-view results under the same or smaller Gaussian budgets.
- Global multi-view attention is too costly at fine patch resolutions.
- View-wise grouping covers neighboring views; 3D grouping reconnects related tokens.
- Rendering error guides recovery beyond one-pass feed-forward prediction.
- Compact generation preserves fine features without extra primitives.
From the original article
1 min read at huggingface.co
Feed-forward 3D Gaussian Splatting (3DGS) promises instant scene reconstruction with high-quality novel-view synthesis. Yet, fine-grained geometry and appearance remain difficult to recover, motivating us to revisit how feed-forward models represent, aggregate, and refine detailed visual information. We introduce FineSplat, a feed-forward 3DGS framework built around fine-grained image patches.
By Hyebin Kim, Hwasik Jeong, Joonhyuk Park · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.