Reproduce “GradRig: Differentiable Weights for Skinned Gaussian Splat Deformation” 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
Nina Vesseron, Élie Michel
GradRig deforms 3D Gaussian splat scenes with coarse rigs by estimating how skinning weights change across space and using those gradients to stretch each splat, without mesh connectivity or neighboring-splat lookups. This reduces visible holes from ordinary per-splat skinning while retaining real-time rendering in the reported examples. Its first-order local approximation can overstretch splats; optional rest-space resampling reduces those artifacts.
Key insights
- Incorporating the spatial gradient of skinning weights into a splat's covariance transformation is crucial for elastic deformation, preventing holes and maintaining visual fidelity under non-rigid motion.
- Analytical computation of skinning weight gradients during rig authoring provides more accurate and efficient deformation information compared to estimating gradients from neighboring splats.
- A dictionary-based adaptive resampling strategy can effectively address inherent limitations of affine splat deformation, allowing for higher visual quality without scene-specific retraining.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.