TokenGS: Decoupling 3D Gaussian Prediction from Pixels with Learnable Tokens
This work addresses the bottleneck of regressing 3D Gaussians from pixels in feed-forward reconstruction, offering a more flexible and robust approach for both static and dynamic scene representation.
TokenGS introduces a novel encoder-decoder architecture with learnable tokens for feed-forward 3D Gaussian Splatting prediction, directly regressing 3D mean coordinates via self-supervised rendering loss. It achieves state-of-the-art reconstruction on static and dynamic scenes, with improved robustness to pose noise and multiview inconsistencies.
In this work, we revisit several key design choices of modern Transformer-based approaches for feed-forward 3D Gaussian Splatting (3DGS) prediction. We argue that the common practice of regressing Gaussian means as depths along camera rays is suboptimal, and instead propose to directly regress 3D mean coordinates using only a self-supervised rendering loss. This formulation allows us to move from the standard encoder-only design to an encoder-decoder architecture with learnable Gaussian tokens, thereby unbinding the number of predicted primitives from input image resolution and number of views. Our resulting method, TokenGS, demonstrates improved robustness to pose noise and multiview inconsistencies, while naturally supporting efficient test-time optimization in token space without degrading learned priors. TokenGS achieves state-of-the-art feed-forward reconstruction performance on both static and dynamic scenes, producing more regularized geometry and more balanced 3DGS distribution, while seamlessly recovering emergent scene attributes such as static-dynamic decomposition and scene flow.