InfiniSplat v2 improves 3D scene rendering from any camera angle
InfiniSplat v2: Analyzing and Improving Feed-Forward 3DGS from Any Views
Computer Vision and Pattern Recognition
Summary
Rendering 3D scenes from pictures taken at different angles is hard, especially when you want the view to change a lot. The authors studied why it’s difficult to learn how to color and shade the scene well and made a system called InfiniSplat v2. It keeps the 3D shape predictions fixed when learning colors and uses known camera positions to better train the system. This way, the system creates clearer and more consistent images when looking at the scene from new angles.
What this means in practice
- •For augmented reality developers: Create realistic 3D scenes that maintain quality when viewed from any angle using feed-forward methods without slow per-scene optimization.
- •For computer graphics engineers: Integrate improved Gaussian attribute learning into rendering pipelines to generate high-quality novel views with large viewpoint changes.
Authors
Yongzhen Hu, Hao Yu, Jiawei Wang, Xianming Zeng, Jianyun Xu, Xingliang Liu, Ruizhen Hu, Xiaowei Zhou, Hujun Bao, Sida Peng
Abstract
Feed-forward Gaussian reconstruction builds 3D scenes directly from input images for novel view rendering, but learning Gaussian attributes from predicted geometry for high-quality rendering under large viewpoint changes remains challenging. We analyze limitations in Gaussian attribute learning and introduce InfiniSplat v2, which improves rendering quality through architectural and training changes. Our method decouples geometry and attribute prediction with a frozen geometry branch, aligns predicted Gaussians using ground-truth camera poses during training, and adopts large-angle target view supervision. Across six datasets, the feed-forward outputs of the same trained model achieve state-of-the-art mean LPIPS and DISTS among the evaluated methods in both pose-conditioned and pose-free settings. In the pose-conditioned setting, mean LPIPS and DISTS decrease by 25.8% and 32.9%, respectively, compared with the strongest evaluated baselines. We further find that opacity learned with only image reconstruction supervision suppresses geometrically inconsistent points. Project page: https://zju3dv.github.io/InfiniSplat2/