AsySplat: Efficient Asymmetric 3D Gaussian Splatting for Long-Sequence Scene Modeling

AsySplat addresses redundant computation in generalizable 3D Gaussian Splatting for long-sequence novel view synthesis by decoupling geometry and appearance modeling. Its geometry branch uses coarse-grained tokens and most of the parameters for multi-view reconstruction, while a smaller appearance branch processes fine-grained tokens to capture visual details. Bilateral connections let the branches guide each other. According to the abstract, on 32-view 960P inputs, AsySplat matches optimization-based methods while achieving nearly 800x speedup, and exceeds the zero-shot performance of state-of-the-art generalizable models with fewer parameters and lower training and inference overhead.