ATSplat: Compact Feed-forward 3D Gaussian Splatting with Adaptive Token Expansion
ATSplat cuts redundant Gaussian counts by using adaptive 3D tokens instead of pixel-aligned primitives.
The paper says existing feed-forward 3DGS methods tie primitive placement to image grids, which can make representations dense and wasteful. ATSplat builds a sparse scaffold from patch-level depth and camera cues, then expands only higher-uncertainty tokens guided by rendering-error supervision. On RealEstate10K and DL3DV, the authors report state-of-the-art rendering quality with more than a 5.7x reduction in Gaussians versus dense feed-forward 3DGS methods. They also report sub-second reconstruction from 12 input images on a single commercial GPU and 1136 FPS rendering at 512 x 960 with 311K Gaussians. HF Daily Papers' note
The paper says existing feed-forward 3DGS methods tie primitive placement to image grids, which can make representations dense and wasteful. ATSplat builds a sparse scaffold from patch-level depth and camera cues, then expands only higher-uncertainty tokens guided by rendering-error supervision. On RealEstate10K and DL3DV, the authors report state-of-the-art rendering quality with more than a 5.7x reduction in Gaussians versus dense feed-forward 3DGS methods. They also report sub-second reconstruction from 12 input images on a single commercial GPU and 1136 FPS rendering at 512 x 960 with 311K Gaussians. HF Daily Papers' note
score 4