Post-Hoc Entropy Coding for Anchor-Based 3D Gaussian Splatting with an Optional Joint-Training Path

Abstract Anchor-based 3D Gaussian splatting enables real-time rendering, but trained scenes remain too large for phones, headsets, and edge servers. Existing post-hoc converters are drop-in yet saturate early, while joint-training codecs compress deeper only by retraining from scratch. We present a flexible compression framework whose primary mode is a strictly post-hoc learned entropy codec for anchor-based Gaussian splatting. The codec treats any trained model as a read-only artifact and learns a per-scene binary hash-grid entropy context with per-channel Laplace distributions. Attributes are coded into a Morton-ordered, progressively decodable bitstream, and the codec requires neither training data nor pipeline access. When the training pipeline is available, the same bitstream format optionally supports a joint-training variant, whose quantization simulation, rate–distortion optimization, and hard straight-through anchor mask align training with deployment. Across three benchmarks and thirteen scenes, the post-hoc mode substantially reduces storage relative to float16 and codebook baselines while preserving quality. It also transfers unchanged to third-party checkpoints and adds no rendering overhead under progressive decoding. The joint mode further increases compression at a modest quality cost and serves as a controlled probe of what retraining adds, though dedicated joint codecs remain stronger in the deepest regime. The framework thus turns post-hoc and joint compression into selectable rungs of one deployable format and decoder.