Reproduce “Trimming the Fat: Efficient Compression of 3D Gaussian Splats through Pruning” 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
Muhammad Salman Ali, Maryam Qamar, Sung-Ho Bae, Enzo Tartaglione
A gradient-informed, iterative pruning method reduces the size of pretrained 3D Gaussian Splatting models without changing their underlying representation. It removes up to about half of the Gaussian primitives with comparable or improved image quality, and can reach roughly 24× standalone compression or 40–56× when combined with an end-to-end coding pipeline.
Key insights
- Gradient information complements opacity: low-opacity Gaussians may remain important when optimization gradients show that they affect image quality.
- Gradual pruning followed by retraining is more effective than deleting the same number of primitives in one operation.
- Successful compression depends on the learned structure and parameters of a dense, pretrained 3DGS model; sparse reinitialization or random feature reinitialization does not reproduce the performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.