SplatWeaver dynamically routes Gaussian primitives for sharper novel view synthesis

SplatWeaver, a generalizable novel view synthesis framework by Harbin Institute of Technology researchers Yecong Wan, Fan Li, Mingwen Shao, and Wangmeng Zuo, learns to allocate 3D Gaussian primitives adaptively across a scene. Most feed-forward Gaussian Splatting methods assign a fixed number of primitives to every pixel or voxel, wasting capacity in smooth regions while starving complex geometry and high-frequency textures. SplatWeaver replaces uniform allocation with a pixel-level routing scheme that coordinates cardinality Gaussian experts—each expert produces a specific number of primitives from 0 to M—letting the model assign more Gaussians where detail demands it and fewer where surfaces are flat.

The routing process is stabilized by a high-frequency prior and attendant guidance module that steers expert selection toward textured regions, plus routing regularization to prevent redundant primitive assignments. By leveraging structural cues, the system concentrates Gaussian capacity on fine structures and complex geometry while suppressing primitives in smooth areas. The result is a more expressive yet compact 3D scene representation that renders unseen views from uncalibrated input images without per-scene optimization. Experiments across diverse scenarios show SplatWeaver consistently outperforms state-of-the-art methods, delivering more faithful novel-view renderings with fewer Gaussian primitives than fixed-allocation baselines.