3D Gaussian Splatting (3DGS) produces high-quality real-time novel views but demands large storage and bandwidth. A common workaround is to map Gaussian attributes to 2D representations and then compress them with mature image codecs. However, the authors of a new arXiv preprint argue that this pipeline leaves rate-distortion performance on the table: representations that render equally well can have very different coding costs, yet they are optimized as if they were equivalent.

Their proposed method, RDCompGS, integrates a differentiable estimate of the coding rate directly into the rate-distortion optimization alongside rendering distortion. A rate constraint adjustment model then tunes the trade-off between bitrate and quality for different compression settings. This lets the optimizer favor Gaussian representations that are cheaper to encode without sacrificing visual fidelity.

Experiments on three benchmark datasets show average Bjøntegaard Delta rate reductions of 37.84% with H.265/HEVC and 49.31% with JPEG XL, meaning the same rendering quality is achieved at substantially lower bitrates. The authors also report that their estimated rate strongly correlates with the actual bitrate under various compression conditions, supporting the usefulness of the proxy in optimization. The paper has been submitted to IEEE Transactions on Multimedia.