Towards Practical Compression of 3D Gaussian Splatting: COSA-GS

Forum topic · 小凯 · 2026-09-28

Summary

3D Gaussian Splatting (3DGS) delivers high-quality novel-view synthesis but demands substantial storage. Existing 3DGS compression methods typically rely on spatial context modeling over irregular 3D representations, which increases training and coding complexity, and floating-point context inference can introduce numerical inconsistencies across platforms that break entropy decoding. The paper proposes COSA-GS, a practical compression framework that builds context via anchor-wise causal factorization without spatial aggregation. Geometry context derived from each anchor's coordinates models a compact learnable anchor latent, which is fused with the geometry context to form an anchor context for attribute coding; the context model uses only linear transformations and activations. Training combines rate-distortion optimization with adaptive Gaussian pruning, and quantization-aware training plus integer inference achieve bit-exact, cross-platform entropy decoding. Experiments show state-of-the-art compression performance with fast, consistent decoding. Code: https://github.com/pengpeng-yu/COSA-GS (arXiv: 2609.30245).

Paper Overview

Field: Computer Vision (CV) Authors: Pengpeng Yu, Yueru Chen, Fei Song, Tai Qin, Qi Zhang, Jing Wang, Yulan Guo Published: 2026-09-24 arXiv: 2609.30245

Abstract (translated)

3D Gaussian Splatting (3DGS) enables high-quality novel-view synthesis but requires substantial storage. Existing compression methods often rely on spatial context modeling over irregular 3D representations, increasing the complexity of training and coding. Meanwhile, floating-point context inference can introduce numerical inconsistencies across platforms, causing entropy-decoding failures.

To address these practical challenges, the authors propose COSA-GS, which constructs context without spatial aggregation through anchor-wise causal factorization. Specifically, geometry context derived from each anchor's coordinates is used to model a compact learnable anchor latent. The anchor latent is then fused with the geometry context to form an anchor context for attribute coding. The resulting context model consists only of linear transformations and activation functions, giving it a simple architecture.

COSA-GS is trained with rate-distortion optimization combined with adaptive Gaussian pruning. In addition, quantization-aware training and integer inference are developed for the context model, enabling bit-exact consistency of entropy-decoding symbols across platforms.

Experiments show that COSA-GS achieves state-of-the-art compression performance while maintaining fast and consistent cross-platform decoding, providing a simple yet effective framework for practical 3DGS compression.

Key Contributions

Anchor-wise causal factorization that avoids costly spatial aggregation over irregular 3D representations
A lightweight context model built solely from linear layers and activation functions
Rate-distortion optimization with adaptive Gaussian pruning for training
Quantization-aware training and integer inference for bit-exact, cross-platform entropy decoding

Code

https://github.com/pengpeng-yu/COSA-GS

---

*Auto-collected on 2026-09-28.*