StereoSplat+ generates 3D scenes from a single stereo pair using diffusion

Most 3D Gaussian Splatting pipelines require multiple views or non-causal access to future frames, limiting use in on-device robotics and AR where only a single stereo rig is available. StereoSplat+ addresses this with a two-component architecture. First, StereoSplat is an input-invariant feed-forward 3D Gaussian estimator that takes a variable number of posed stereo pairs and predicts high-quality Gaussians by fusing geometry cues from a cost-volume branch and a triplane-based 3D volume branch, using continuous pose encoding to generalize across view counts.

Since multiple stereo pairs are typically unavailable at inference, the second component introduces a diffusion-enhanced progressive inference scheme. Starting from one stereo pair, the system renders novel stereo views from the predicted 3DGS, refines them with a one-step diffusion enhancer, and feeds them back as additional inputs to update the 3DGS. Experiments on KITTI-360 show significant improvements in novel-view rendering quality and geometry accuracy, especially in occluded regions and under strong view extrapolation, outperforming recent feed-forward 3DGS baselines. Accepted as a conference paper at IROS2026.