InfiniSplat v2: Analyzing and Improving Feed-Forward 3DGS from Any Views
Yongzhen Hu, Hao Yu, Jiawei Wang, Xianming Zeng, Jianyun Xu, Xingliang Liu, Ruizhen Hu, Xiaowei Zhou and 2 others
Zhejiang University · Hello TransTech · Shenzhen University · Causa Robotics
Published October 14, 2026 · on the feed October 16, 2026 · ▲ 122 · arXiv 2610.20341
The 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/
The Brief
2:04A short podcast summary of the paper. Written and voiced by AI.
Transcript
InfiniSplat v2. Analyzing and Improving Feed-Forward 3DGS from Any Views. By Yongzhen Hu and nine others, from Zhejiang University, Hello TransTech, and two other institutions. Published on the fourteenth of October. Its code repository is worked out: about four hundred source files, with a readme and a licence.
Feed-forward Gaussian reconstruction builds a 3D scene straight from input images, ready to render from new viewpoints. The hard part is learning the attributes of each Gaussian from predicted geometry well enough to render at high quality when the viewpoint changes a lot. Geometry may be predicted reasonably, yet the attributes attached to it still need to hold up when the scene is seen from far away from the inputs, and that is where rendering quality suffers.
The authors analyze where Gaussian attribute learning falls short, and answer with InfiniSplat v2, a set of changes to both the architecture and the training.
Three changes are described. Geometry and attribute prediction are decoupled, with the geometry branch frozen. During training, the predicted Gaussians are aligned using ground-truth camera poses. And supervision comes from target views at large angles. That last change speaks directly to the problem as stated: if quality suffers under large viewpoint changes, the training now asks the model to render exactly those views.
Across six datasets, outputs from one and the same trained model reach state of the art mean LPIPS and DISTS among the evaluated methods, in both the pose-conditioned and the pose-free setting. With poses given, mean LPIPS falls by twenty-five point eight percent and mean DISTS by thirty-two point nine percent against the strongest evaluated baselines. A further observation: opacity learned only from image reconstruction supervision suppresses points that are geometrically inconsistent.
A single trained model covers both settings, with poses and without, and leads the evaluated methods in each. Alongside the scores comes that small finding about opacity. Trained only to reproduce images, it ends up filtering out geometry that does not fit, a cleanup no one asked it to perform.
More with an account
Listen to every day’s papers
Daily playlists, deep dives, quests for the papers, and quizzes on their key terms.
Join free