Trained Gaussian-splat scenes no longer need retraining to shrink: post-hoc entropy coding reaches 6.0x on 13 scenes at a 0.03 dB quality change

Compression no longer requires retraining from scratch: an already-trained anchor-based Gaussian-splat model whose training pipeline is unavailable can be cut to roughly one sixth of its size with almost no change in rendered quality. Earlier post-hoc converters mapped weights to float16, scalar grids or codebooks under rules identical for every scene, so the redundancy they exploited was frozen when training ended, ratios saturated early and distortion rose sharply at low bit budgets. Across 13 scenes from Mip-NeRF 360, Tanks and Temples and Deep Blending, the post-hoc mode averages 6.0x compression at a mean quality change of -0.03 dB; on the same scenes the joint rung reaches 12.9x at -0.33 dB.

The encoder reads only the model file, needing no training images, distillation recipe or pipeline access, and fits one scene in about nine minutes on a single GPU. Joint-training codecs put quantization simulation and entropy contexts inside the optimization loop, retrain every scene from scratch, and the deployed model is no longer the validated one. Four checkpoints the team did not train, two native Scaffold-GS and two Octree-GS, compress with no adaptation at a mean 6.4x, with a paired quality change of -0.049 dB against the original checkpoints.

The bitstream is split into independently decodable spatial chunks, so a client renders each region as its chunk arrives instead of waiting for the whole file. Fixed-rule baselines require a full download before any rendering can begin. On bonsai and bicycle, a quarter of the stream already yields a recognizable image and quality climbs with cumulative received bytes; progressive decoding adds no rendering overhead.

When the training pipeline is available, the same bitstream format adds a rung: quantization simulation, a rate-distortion loss and a hard straight-through anchor mask during training roughly double the ratio again. The post-hoc mode is bounded by redundancy left behind at training time, and below 7 bits the frozen symbol statistics become the bottleneck that a re-fit cannot recover. On bonsai the post-hoc mode holds 30.61 dB at 6 bits while the joint-plus-mask rung holds 31.13 dB at the same depth, a 0.52 dB gap; the 13-scene mean is 8.8 MB, 12.9x, -0.33 dB, keeping 76% of anchors on average.

The demonstrated setting is anchor-based Gaussian splatting, represented by Scaffold-GS, where each anchor stores a feature, offsets and scalings that lightweight MLPs decode into neural Gaussians; evaluation covers nine Mip-NeRF 360 scenes, two from Tanks and Temples and two from Deep Blending, all on one RTX 4090D with test views rendered at 1600 px. The post-hoc rung targets deployers who hold only the model file, while the joint rung targets teams that still own the training pipeline.

The post-hoc rung is bounded by the redundancy the model's training left behind, and below 7 bits the frozen symbol statistics become the bottleneck that a re-fit cannot recover. Lifting that bound with the joint rung requires retraining from scratch, and dedicated joint codecs remain stronger in the deepest regime. Quantization currently uses one grid per channel, with per-anchor adaptive steps and offset-level masks named as next steps.