FineSplat (ours)

YoNoSplat

2Xplat

ZipSplat

Results on DL3DV at 224×224 resolution with 24 context views in the pose-free setting.

Overview

Fine Patches

FineSplat works with fine-grained image patches, recovering substantially finer visual details while staying computationally efficient.

Local Groups

Fine patches give Stage 1 a strong initial scene. Stage 2 then further enhances its fine structure and appearance with error feedback and 3D grouping: the Stage 1 tokens are lifted into 3D, where tokens that lie close to each other interact, guided by the rendering errors of the input views, and grow into richer, more expressive Gaussian tokens.

Input views of the scene in the viewer.

Detailed Splats

FineSplat uses far fewer Gaussians than pixel-aligned methods and still reconstructs more detailed scenes than existing methods. Because each Gaussian comes from a small patch, it stays small and local, so thin structures, sharp edges and fine textures such as signboard lettering come through, where other methods blur them or spend many more primitives.

Quantitative Results

Benchmark performance

Novel view synthesis on the DL3DV benchmark, evaluated in both posed and pose-free settings with a varying number of input views.

Citation

@article{kim2026finesplat,
  title   = {FineSplat: Fine Patches, Local Groups, Detailed Splats},
  author  = {Kim, Hyebin and Jeong, Hwasik and Park, Joonhyuk and Kang, Gyeongjin and
             Kim, Jungwoo and Kim, Younghyun and Park, Eunbyung},
  journal = {arXiv preprint arXiv:2610.25011},
  year    = {2026}
}