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

TL;DR

ATSplat introduces a feed-forward 3D Gaussian Splatting method that uses adaptive 3D tokens to enable scene-aware primitive placement, reducing redundancy and improving efficiency. It lifts coarse depth and camera cues into sparse anchor tokens, expands uncertain ones adaptively, and achieves high-quality novel-view synthesis with significantly fewer Gaussians.

Why It Matters

This approach enables fast, compact 3D scene reconstruction and rendering suitable for real-time applications on consumer hardware, without sacrificing visual fidelity. It bridges the gap between optimization-based 3DGS and feed-forward methods by restoring adaptive capacity allocation.

Key Findings

• ATSplat reduces Gaussian count by over 5.7 times compared to dense feed-forward 3DGS methods while maintaining superior rendering quality.
• The adaptive token expansion mechanism dynamically allocates primitives based on scene complexity, improving representation efficiency.
• Reconstruction completes in under a second and rendering runs at 1136 FPS on a single GPU using only 311K Gaussians from 12 input views.