StereoSplat+: Feed-Forward Stereo Gaussian Splatting with Diffusion-Assisted Progressive Inference

Published in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2026), 2026

StereoSplat+ reconstructs a 3D scene from a single stereo pair. It renders additional stereo views from an initial 3D Gaussian representation, refines those views with a one-step diffusion model, and uses them to update the reconstruction. On KITTI-360, the method improves novel-view rendering and geometry, especially in occluded regions and for viewpoints far from the input cameras.

arXiv preprint · Project · Code · Demo

StereoSplat+ renders new views from the initial reconstruction, refines them with diffusion, and feeds them back to update the 3D Gaussians
StereoSplat+ inference pipeline, Figure 1 by Zihua Liu and Masatoshi Okutomi · CC BY-NC-ND 4.0