Our paper “POTR: Post-Training 3DGS Compression” has been accepted for publication in the IEEE Transactions on Circuits and Systems for Video Technology (TCSVT).
POTR is a post-training compression codec designed to address the large storage footprint of 3D Gaussian Splatting (3DGS). Unlike many existing approaches, POTR compresses already-trained 3DGS models without requiring access to the original training pipeline or fine-tuning.
Overview of the POTR pipeline, combining pruning, spherical harmonics energy compaction, quantization, and entropy compression.
Key Contributions
- Efficient Pruning: Instead of relying on simple heuristics, POTR calculates exactly what the quality impact is of removing each individual splat. This strategy allows the codec to use 2–4× fewer splats, leading to 1.5–2× faster rendering.
- SH Energy Compaction: Since lighting coefficients typically account for over 80% of the model size, POTR introduces an optimization approach to compress these coefficients into a lower-entropy representation. This increases sparsity in the AC coefficients to around 97% without requiring retraining.
For the full evaluation, qualitative results, and implementation details, please visit the project page
