Reproduce “GrainGS: Gradient-Decoupled Gaussian Splatting for Efficient Dynamic Novel View Synthesis” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jiahao He, Yihua Shao, Zhengkai Zhao, Pan Gao, Fei Ma, Jingcai Guo, Hao Tang, Nicu Sebe and 1 more

GrainGS, a dynamic 3D Gaussian Splatting framework, achieves robust, high-fidelity novel view synthesis for dynamic scenes by decoupling canonical geometry, temporal deformation, and dynamic appearance through a gradient-isolated optimization. This approach yields an average 36.98 dB PSNR on D-NeRF, up to 435.6 FPS rendering speed, and a compact model size of 4.67 MB, surpassing prior methods.

Key insights

  • Explicitly isolating canonical geometry from deformation gradients through a stop-gradient mechanism is essential for preserving the stability and integrity of the time-invariant scene representation.
  • Combining a hierarchical anchor scaffold for structural compactness with independent per-Gaussian temporal offsets allows for both global structural regularity and expressive fine-grained local motion.
  • Decomposing appearance into canonical and residual components prevents dynamic photometric changes from corrupting the underlying geometric and deformation parameters, enhancing overall fidelity.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.