AsySplat: Efficient Asymmetric 3D Gaussian Splatting for Long-Sequence Scene Modeling
AsySplat splits geometry and appearance work to cut redundant computation in long-sequence 3D Gaussian Splatting.
The paper argues that high-precision geometry is not always necessary for strong novel view synthesis, while appearance details are easier to learn. Its architecture sends coarse tokens through a larger geometry branch and fine tokens through a smaller appearance branch, with bilateral connections between them. On 32-view 960P inputs, the authors say it matches optimization-based methods with nearly an 800x speedup and beats state-of-the-art generalizable models in zero-shot performance using fewer parameters. HF Daily Papers' note
The paper argues that high-precision geometry is not always necessary for strong novel view synthesis, while appearance details are easier to learn. Its architecture sends coarse tokens through a larger geometry branch and fine tokens through a smaller appearance branch, with bilateral connections between them. On 32-view 960P inputs, the authors say it matches optimization-based methods with nearly an 800x speedup and beats state-of-the-art generalizable models in zero-shot performance using fewer parameters. HF Daily Papers' note
score 4