StoSplat: Ray-Aligned Stochastic Preconditioning for Feed-Forward 3D Gaussian Splatting

Feed-forward 3D Gaussian splatting enables single-pass reconstruction from multi-view images, but widely-applied voxel-aligned pipelines usually produce sharp and anisotropically ill-conditioned Gaussian centers, leading to floaters and unstable geometry in sparse-view settings. To remedy this, we introduce StoSplat, a training-time stochastic preconditioning framework that perturbs predicted Gaussian centers with annealed, ray-aligned anisotropic noise. This perturbation smooths the expected rendering objective in the geometric variable where instability arises, while leaving covariance, opacity, color, and the inference architecture unchanged. StoSplat introduces negligible additional computation during training without any computational overhead during inference. Experiments on widely used benchmarks including RealEstate10K, ScanNet, and ACID demonstrate that StoSplat achieves state-of-the-art performance, while producing more stable geometry with fewer floaters and boundary artifacts, which demonstrates the effectiveness of our method in improving the performance of existing feed-forward 3D reconstruction without architectural changes.