G3Splat: Geometrically Consistent Generalizable Gaussian Splatting
It addresses the challenge of geometrically meaningful 3D scene representation for pose-free generalizable splatting, which is incremental by building on prior multi-view prediction networks.
The paper tackles the problem of learning geometrically consistent 3D Gaussian splats from images under self-supervision, showing that view-synthesis loss alone is insufficient. It introduces G3Splat, which enforces geometric priors to achieve state-of-the-art performance in reconstruction, pose estimation, and novel-view synthesis on RE10K and strong zero-shot generalization on ScanNet.
3D Gaussians have recently emerged as an effective scene representation for real-time splatting and accurate novel-view synthesis, motivating several works to adapt multi-view structure prediction networks to regress per-pixel 3D Gaussians from images. However, most prior work extends these networks to predict additional Gaussian parameters -- orientation, scale, opacity, and appearance -- while relying almost exclusively on view-synthesis supervision. We show that a view-synthesis loss alone is insufficient to recover geometrically meaningful splats in this setting. We analyze and address the ambiguities of learning 3D Gaussian splats under self-supervision for pose-free generalizable splatting, and introduce G3Splat, which enforces geometric priors to obtain geometrically consistent 3D scene representations. Trained on RE10K, our approach achieves state-of-the-art performance in (i) geometrically consistent reconstruction, (ii) relative pose estimation, and (iii) novel-view synthesis. We further demonstrate strong zero-shot generalization on ScanNet, substantially outperforming prior work in both geometry recovery and relative pose estimation. Code and pretrained models are released on our project page (https://m80hz.github.io/g3splat/).