RacerGS: Residual-Autoregressive Context for 3D Gaussian Splatting Compression

RacerGS compresses 3D Gaussian Splatting scenes using a hash projection model, a FiLM-modulated autoregressive GRU context model, and residual entropy coding of anchor attributes. The authors report over 150x compression relative to vanilla 3DGS and 28x relative to Scaffold-GS, with comparable rendering quality across five datasets.