ATSplat: Compact Feed-Forward 3D Gaussian Splatting with Adaptive 3D Tokens

Paper overview

• Research field: Computer Vision
• Authors: Cho In, Jeonghwan Cho, Mijin Yoo
• arXiv: 2507.18389

Abstract

3D Gaussian Splatting (3DGS) achieves high-quality novel-view synthesis by optimizing freely placed primitives in 3D and adaptively densifying them in under-reconstructed regions. However, this scene-adaptive capacity allocation is largely missing from existing feed-forward 3DGS methods. These methods commonly regress Gaussians at input pixels and lift them along camera rays.

Such pixel-aligned formulations make the number and placement of primitives dependent on image resolution and input viewpoints rather than scene complexity. This results in dense and often redundant Gaussian sets.

ATSplat is a feed-forward 3DGS framework designed to restore the adaptive allocation capability of 3DGS optimization through Adaptive 3D Tokens.

Method

ATSplat first lifts coarse patch-level depth and camera cues into sparse 3D anchor tokens, which form a compact scaffold for the scene. Each token is then regressed as a local Gaussian with a learnable 3D offset. This formulation decouples primitive placement from the input image grid.

The adaptive token expansion module predicts token-level uncertainty scores. These scores are supervised by a rendering-error map, allowing the model to identify areas where additional capacity is needed. A learnable expansion layer then selectively expands high-uncertainty tokens.

The resulting sparse-to-adaptive design allows ATSplat to concentrate primitives in challenging regions while maintaining a compact representation.

Experimental results

Experiments on two representative datasets, RealEstate10K and DL3DV, demonstrate that ATSplat:

• Achieves state-of-the-art rendering quality.
• Reduces the number of Gaussians by more than 5.7 times compared with dense feed-forward 3DGS methods.
• Reconstructs scenes from 12 input images at 512×960 resolution in less than one second on a single commercial GPU.
• Renders high-quality novel views at 1136 FPS at 512×960 resolution using only 311K Gaussians.

The results show that adaptive 3D token allocation can provide efficient, compact, and high-performance feed-forward novel-view synthesis.