ATSplat: Compact Feed-forward 3D Gaussian Splatting with Adaptive Token Expansion

ATSplat introduces Adaptive 3D Tokens for feed-forward 3D Gaussian Splatting. It lifts coarse patch-level depth and camera cues into sparse 3D anchor tokens, then predicts local Gaussians with learnable 3D offsets. A token-level uncertainty score, supervised by rendering error maps, selectively expands difficult regions. This decouples primitive allocation from input image grids and viewpoints. On RealEstate10K and DL3DV, the authors report state-of-the-art rendering quality with more than 5.7x fewer Gaussians than dense feed-forward 3DGS methods. Using 12 512×960 input images, reconstruction takes under one second on a single commercial GPU, while rendering reaches 1136 FPS with 311K Gaussians.