FineSplat improves 3D scene reconstruction with detailed patches
FineSplat: Fine Patches, Local Groups, Detailed Splats
Computer Vision and Pattern Recognition
Summary
Reconstructing 3D scenes quickly while keeping fine details is hard. The authors present FineSplat, a method that breaks images into small patches and groups visual information in two ways: by nearby views and by 3D space connections. This approach helps recover detailed shapes and appearances more accurately. They also use error feedback to refine the results and keep the model efficient by generating compact representations.
What this means in practice
- •For augmented reality developers: Create detailed 3D models of real-world scenes instantly for AR experiences using FineSplat’s efficient reconstruction.
- •For 3d content creators: Produce higher quality 3D scene captures with fine geometry and appearance details without increased computational costs.
Authors
Hyebin Kim, Hwasik Jeong, Joonhyuk Park, Gyeongjin Kang, Jungwoo Kim, Younghyun Kim, Eunbyung Park
Abstract
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. Since global multi-view attention becomes prohibitively expensive at such high token resolutions, we develop a dual-grouping strategy that combines two complementary forms of locality: view-wise grouping enables efficient interaction among neighboring views, while 3D spatial grouping reconnects geometrically related tokens that may never have interacted under view-wise grouping by lifting them into the 3D scene space. We further leverage reconstruction error as direct feedback, providing an explicit rendering-based signal to recover fine-grained structures and appearance, thereby overcoming the limitations of one-pass feed-forward prediction. Finally, a compact Gaussian generation scheme preserves the benefits of fine-grained features without unnecessarily increasing the number of primitives. Together, these designs enable FineSplat to recover richer geometry and appearance while remaining efficient and compact, achieving state-of-the-art results by a substantial margin, under the same or smaller Gaussian budgets than existing methods.