Feed-forward 3D Gaussian Splatting methods aim to reconstruct a scene directly from input images, but they often struggle when the target viewpoint differs sharply from the observed views. In InfiniSplat v2, Hu et al. analyze where this fails and propose architectural and training changes that keep geometry prediction separate from the prediction of Gaussian attributes.

Key changes include a frozen geometry branch, so attribute prediction cannot distort the learned geometry; aligning predicted Gaussians with ground-truth camera poses during training; and supervising with large-angle target views. The paper also finds that opacity learned under image-reconstruction-only supervision tends to suppress geometrically inconsistent points, offering a possible explanation for the improvement.

Across six datasets, the same trained model achieves state-of-the-art mean LPIPS and DISTS among evaluated methods in both pose-conditioned and pose-free settings. In the pose-conditioned setting, the authors report mean LPIPS and DISTS decreases of 25.8% and 32.9%, respectively, compared with the strongest evaluated baselines.