D3DGS-Bench: a benchmark for dynamic 3D Gaussian splatting compression
New York University
D3DGS-Bench is an open benchmark for the compression of dynamic 3D Gaussian splatting (3DGS) sequences. A dynamic 3DGS sequence stores one Gaussian set per frame, and how consecutive sets relate is determined by the estimator that produced them: tracked estimators move a fixed set of Gaussians, semi-tracked estimators update a persistent set while adding and removing Gaussians, and untracked estimators reconstruct every frame independently. Because this relation decides what a codec can exploit, D3DGS-Bench evaluates every codec across estimators rather than on the output of a single one. The benchmark comprises 24 input sequences (four estimators, six N3DV scenes, 300 frames each), nine codecs behind a common interface (six static GS codecs, each applied per frame and to concatenated groups of frames, and three dynamic GS codecs), and a rendering-based evaluation against two references: the uncompressed input frames and the held-out camera. On this benchmark, concatenating 30 frames before static coding reduces the rate by 42–72% on tracked sequences but by at most 3.5% on untracked ones. All rate–distortion points, BD-rates, bitstream compositions, runtimes and memory footprints are released with full provenance, together with scripts that reproduce the complete experiment grid.
What the data is, how the two coding branches differ, and what the numbers on this page mean.
3D Gaussian splatting represents a scene as a set of 3D Gaussians, each with a position, a scale and rotation, an opacity and spherical-harmonic colour coefficients. A dynamic sequence stores one such set per frame: here 300 frames of each of the six N3DV scenes, 118 k to 845 k Gaussians per frame.
The estimator is the method that produced the sequence from multi-view video. Tracked estimators move a fixed set of Gaussians (3DGStream, 4D Gaussians), semi-tracked ones update a persistent set while adding and removing Gaussians (QUEEN), and untracked ones reconstruct every frame independently (INRIA 3DGS). That relation is what a codec can exploit, so every codec is run on all of them.
In the per-frame branch each frame is coded on its own. In the concatenated branch the Gaussians of a group of L frames (a GoF) are merged into one Gaussian set, every Gaussian carries the index of its frame, and the static codec codes the group as a single set; the decoder splits it back by that index. The codec's own quantization, prediction, ordering and entropy coding then find the redundancy between frames.
Rate is MB per frame, quality is PSNR against a reference, and a pair of curves is summarised by its BD-rate: the average rate difference at equal quality, negative when the first curve spends fewer bits. Every BD-rate on this page compares a codec with itself, concatenated against per-frame, on the same input and the same rate points.
A dynamic codec codes a sequence as key frames (I) and inter-coded frames (P): the I-frames are coded as Gaussian sets, and the P-frames are predicted from them by a learned motion model. D-FCGS is the published method of that kind. D-GPCC is ours: it replaces only D-FCGS's I-frame codec by concatenated G-PCC coding and keeps the P-frame model, its checkpoint and its parameters unchanged, which isolates what the key-frame codec contributes. Its per-frame ablation codes the same I-frames one at a time.
Six-scene average BD-rate of GoF-30 concatenation relative to per-frame coding with the same codec at the same rate points. Negative values indicate bit-rate savings.
How to read it: rows are the six static codecs, columns the three sequence types. Select any cell to load that comparison into the curves below.
All measured rate–distortion curves. The reference selector sets the quality reference: the uncompressed input frames (identical-source evaluation) or the captured images of the held-out camera (ground-truth evaluation).
How to use it: start from one of the shortcuts below the selectors, then tick series in the list on the left to add or remove curves. Under Input frames each estimator is drawn in its own panel, because its PSNR is measured against its own frames; the estimator selector follows the shortcut you chose. Drag on the plot to zoom, double-click to reset, hover a point for its rate and quality. Solid lines are concatenated or inter-coded runs, dashed lines the per-frame branch of the same codec.
One cell per experiment, estimator and scene. Each value is the BD-rate of the experiment relative to its anchor, computed on input-referenced PSNR; hatched cells mark experiments that were not run.
How to use it: columns are estimator × scene, rows the experiments of the benchmark. A dot in the corner marks a cell that also has ground-truth scores. Select any cell to load its curves into the explorer above.
Distribution of each run's bitstream over the streams written by its codec, at the level of detail recorded for that run.
How to use it: choose a codec, then an input and a group length; each bar is one run of that selection, split into the streams its codec writes. Select a bar to see the same run as a ring and as exact byte counts in the table. Display switches between each stream's share of the bitstream and its MB per frame, and the spherical-harmonic switch merges or separates the SH bands.
Each upstream project is retrieved at the commit listed on its card and patched by its setup.sh; these patches constitute the complete set of our modifications.
Each cell shows the cost of predicting frame j from frame i; brighter cells indicate higher similarity. ρ relates the cost between distant frames to that between adjacent frames; values above 1 indicate temporal structure. Flame steak, 300 × 300 frame pairs.
How to read it: a bright band along the diagonal means consecutive frames predict one another well, which is the redundancy a codec can exploit; a uniformly dark map means the frames share little, as in independently reconstructed sequences.
PSNR against renders of the uncompressed input frame from every calibrated camera, pooled over all pixels, views and 300 frames. It isolates the coding loss; curves of different estimators are therefore not on a common scale.
A score is rejected before it reaches a table when not every frame was scored, when pooled RGB PSNR falls below 15 dB, or when the mean of per-frame decibels exceeds the pooled value by 3 dB or more, the signature of a few damaged frames or views.
PSNR, SSIM and LPIPS against the held-out camera. All estimators share a common reference, which allows published results to be shown alongside ours; both branches, however, saturate at the reconstruction quality of the estimator. Ground-truth scores are available for the estimator sweep (G-PCC at GoF 150), GSCodec-S, GSCodec-D, D-FCGS and D-GPCC, and for the published and uncompressed anchors. The remaining static codecs at GoF 30 were scored against the input only and are hidden from the series list under this reference.
Rate is given in MB per frame (total bytes ÷ frames ÷ 106), including all I-frames. G-PCC and HGSC code the frame index of a concatenated group in their bitstreams, and SPZ preserves the input order; L-GSC, LGSCV and GSCodec-S recover it from encoder-side ordering and are charged log230 = 4.91 bits per Gaussian, the cost of a uniformly coded index stream. BD-rate integrates cubic fits of log rate over the quality range shared by the two curves.
The i-th rate point of each scene is averaged over the six scenes, and the BD-rate is computed between the averaged curves.
Every data point is a row of results/tables in the repository (commit ) and can be traced to its raw score file.
D3DGS-Bench accompanies the paper Robust, Estimator-Agnostic Dynamic 3DGS Compression (arXiv:2609.25633). If you use the benchmark, please cite it.
@misc{wang2026robustestimatoragnosticdynamic3dgs,
title={Robust, Estimator-Agnostic Dynamic 3DGS Compression},
author={Chenjunjie Wang and Zixi Huang and Yao Wang and Jona Ballé},
year={2026},
eprint={2609.25633},
archivePrefix={arXiv},
primaryClass={eess.IV},
url={https://arxiv.org/abs/2609.25633},
}