FineSplat: Fine Patches, Local Groups, Detailed Splats

FineSplat. Fine Patches, Local Groups, Detailed Splats. By Hyebin Kim and six others, from Yonsei University. Published on the eighteenth of October. Its code repository is worked out: about seventy-five source files, with a readme and a licence.

Feed-forward 3D Gaussian Splatting promises instant reconstruction of a scene, with high-quality synthesis of new views, in a single pass. What it still finds hard is the fine grain: detailed geometry and detailed appearance. That leads the authors to revisit three things about feed-forward models: how they represent detailed visual information, how they aggregate it, and how they refine it.

FineSplat is built around fine-grained image patches. Smaller patches mean more tokens per image, and that creates the central difficulty, because global attention across all views becomes prohibitively expensive at such high token resolutions. The design that follows is about keeping the fine patches while making them affordable.

The first piece is a dual-grouping strategy that combines two kinds of locality. View-wise grouping lets neighbouring views interact efficiently. But tokens that belong together geometrically may never meet under view-wise grouping, so a second, 3D spatial grouping lifts tokens into the scene space and reconnects the ones that are geometrically related.

The second piece uses reconstruction error as direct feedback. That gives the model an explicit, rendering-based signal for recovering fine structure and appearance, which a single feed-forward pass would otherwise lack.

The third is a compact Gaussian generation scheme. It keeps the benefit of the fine-grained features without unnecessarily multiplying the number of primitives.

Together, these let FineSplat recover richer geometry and appearance while staying efficient and compact. The authors report state-of-the-art results by a substantial margin, and they reach them under the same Gaussian budget as existing methods, or a smaller one. The abstract gives no specific figures.

Each part of FineSplat answers a cost created by the part before it. Fine patches bring too many tokens, so attention is grouped locally. Local groups can miss geometric neighbours, so a second grouping in 3D reconnects them. One pass can miss detail, so the rendering error is fed back. And fine features could inflate the scene, so Gaussian generation is kept compact.