FineSplat: Fine-Grained Patches and Dual Grouping for Detailed View Synthesis

FineSplat: A New Level of Detail in View Synthesis

A team from KAIST led by Hyebin Kim, Hwasik Jeong, Joonhyuk Park, Gyeongjin Kang, Jungwoo Kim, Younghyun Kim, and Eunbyung Park has introduced FineSplat, a feed-forward 3D Gaussian Splatting architecture that reconstructs photorealistic scenes with a level of detail unprecedented for this approach. The method works without camera pose information (pose-free), which is critical for practical applications in scanning and media production.

Three Layers of Architecture

3D Gaussian Splatting promises instant, high-quality reconstruction, but fine-grained geometry and texture remain a hard problem. Global multi-view attention demands prohibitive computational cost at high token resolutions. FineSplat tackles the problem comprehensively across three layers.

The first layer is fine-grained features. The system works with high-resolution image patches instead of global scene features, extracting local features from individual frame regions. This better preserves subtle texture and micro-geometry detail.

The second layer is a dual grouping attention strategy that combines two complementary forms of locality: • View-wise grouping: efficient interaction between geometrically neighboring views without a costly global transformer • 3D spatial grouping: reconnecting tokens that are related in the scene's three-dimensional space but would otherwise fail to interact under a simple view-wise strategy

The third layer uses reconstruction error as an explicit feedback signal. Rendering-based feedback applies direct rendering-based comparison against target images, providing explicit geometric and textural correction. This overcomes the limitations of single-pass feed-forward prediction. A compact Gaussian generation scheme preserves the benefits of fine-grained features without unnecessarily inflating the number of primitive elements.

Results and Scope of Application

FineSplat achieves state-of-the-art results by a significant margin, recovering richer geometry and texture with roughly 4× fewer Gaussians than existing methods. The method is equally effective for posed novel view synthesis and for the pose-free scenario.

The Ecosystem of Similar 2026 Methods

FineSplat fits into a rapidly growing class of feed-forward 3DGS approaches. Closely related work from this year includes: • DeltaSplat — iterative Gaussian refinement for pose-free feed-forward 3D Gaussian Splatting • InfiniSplat v2 — analysis and improvement of feed-forward 3DGS for synthesis from arbitrary views • GrapeSplat — geometry-grounded reconstruction through amalgamated pose-free encoding in 3DGS • AIMS — anchor-integrated multi-view synthesis for scalable novel view rendering • Anchor3DGS — feed-forward 3D Gaussian Splatting with a compact anchor-based representation • Tele360 — real-time human reconstruction from sparse unposed cameras • CoTL — a compact, token-light representation for 3D scenes

The paper was published on October 18, 2026, and has received 36 upvotes on Hugging Face and 32 stars on GitHub. The source code and an interactive demo are available at github.com/hyebinny/FineSplat and hyebinny.github.io/FineSplat/