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

Wang, Wenqing; Zeng, Huimin; Fu, Yun

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

Wenqing Wang, Huimin Zeng, Yun Fu

Northeastern University

NeurIPS 2026

Paper Code (Coming Soon)

RacerGS teaser: bits per anchor of a prior method and RacerGS

Previous methods have limited spatial context, inefficient entropy modeling, and no causal dependencies across anchor feature groups, which leads to high coding costs. RacerGS addresses these limitations through compact context projection, residual coding, and autoregressive GRU refinement, shifting the bitrate distribution toward fewer bits per anchor (bottom) while preserving rendering fidelity.

Abstract

3D Gaussian Splatting (3DGS) has emerged as a promising approach for fast, high-quality novel-view rendering. However, its large parameter size remains a bottleneck for storage and transmission. We propose RacerGS, a compact 3DGS compression framework that achieves scene reconstruction at significantly reduced storage sizes with preserved fidelity. To obtain compact and spatially coherent features, we design a hash projection model that projects the anchor hash encodings into compact context features. Additionally, we introduce a FiLM-modulated autoregressive GRU context model that utilizes causal dependencies among anchor feature groups and uses recurrent hidden states to refine entropy-model parameters, improving bitrate efficiency. Furthermore, we adaptively entropy-code quantized anchor-attribute residuals relative to their context-predicted means, leveraging spatial consistency to reduce symbol entropy.

Overall, RacerGS achieves more than 150× compression over vanilla 3DGS and 28× over Scaffold-GS, while providing comparable rendering quality. Extensive experiments across five datasets demonstrate that RacerGS outperforms prior 3DGS compression methods in both storage reduction and rate-distortion performance.

150×

smaller than vanilla 3DGS

28×

smaller than Scaffold-GS

−26%

size vs. HAC++ on Mip-NeRF360 at the same PSNR

171

FPS rendering on Mip-NeRF360

Method

RacerGS pipeline overview (tap to enlarge)

RacerGS builds on an anchor-based 3DGS codec and compresses each anchor's feature f^a, scaling l, and offsets o with three components. The refined entropy parameters are combined with the spatial prior to form a Gaussian mixture model for residual entropy coding.

(1) Hash projection model

Instead of feeding raw multi-resolution hash encodings f^h to the entropy networks, an MLP projects them into compact, fused context features f^c, so the entropy context stays low-dimensional while the hash grid keeps its capacity.

(2) Adaptive residual quantization

Anchor attributes are quantized and entropy-coded as residuals around their context-predicted means with learned step sizes, r = ⌊(f − μ^s) ⊘ q⌉, which contracts the dynamic range and lowers symbol entropy.

(3) FiLM-modulated autoregressive GRU

Each anchor feature is split into chunks. A GRU advances one step per decoded chunk, its hidden state is FiLM-modulated, and a weight-shared, chunk-positional predictor refines the entropy parameters of the next chunk.

Qualitative Comparisons

Zoomed-in crops with PSNR / size per method. RacerGS preserves edges and material details at a fraction of the storage.

garden (Mip-NeRF360): fewer leaf artifacts

playroom (DeepBlending): clearer book titles

train (Tanks&Temples): sharper metal structures

lego (Synthetic-NeRF)

quebec (BungeeNeRF)

Rate-Distortion Results

RD curves on Synthetic-NeRF, Mip-NeRF360, Tanks&Temples, DeepBlending, and BungeeNeRF. RacerGS reaches a better rate-distortion trade-off than codebook, pruning, entropy, and anchor-based baselines across datasets.

Quantitative Comparisons

Storage sizes in MB. The best and second-best compression results are highlighted in green and blue; lr / hr denote the low- and high-rate settings.

Method	Synthetic-NeRF				Mip-NeRF360				Tanks&Temples
Method	PSNR↑	SSIM↑	LPIPS↓	Size↓	PSNR↑	SSIM↑	LPIPS↓	Size↓	PSNR↑	SSIM↑	LPIPS↓	Size↓
3DGS	33.80	0.970	0.031	68.46	27.46	0.812	0.222	750.9	23.69	0.844	0.178	431.0
Scaffold-GS	33.41	0.966	0.035	19.36	27.50	0.806	0.252	253.9	23.96	0.853	0.177	86.50
Lee et al.	33.33	0.968	0.034	5.54	27.08	0.798	0.247	48.80	23.32	0.831	0.201	39.43
Compressed3D	32.94	0.967	0.033	3.68	26.98	0.801	0.238	28.80	23.32	0.832	0.194	17.28
EAGLES	32.51	0.964	0.039	4.26	27.14	0.809	0.231	58.91	23.28	0.835	0.203	28.99
SizeGS	32.27	0.956	0.043	2.24	27.19	0.788	0.253	17.05	24.11	0.833	0.191	10.51
LightGaussian	32.73	0.965	0.037	7.84	27.00	0.799	0.249	44.54	22.83	0.822	0.242	22.43
SOG	31.37	0.959	0.043	2.00	26.56	0.791	0.241	16.70	23.15	0.828	0.198	9.30
Navaneet et al.	32.99	0.966	0.037	3.10	27.12	0.806	0.240	19.33	23.44	0.838	0.198	12.50
Reduced3DGS	33.02	0.967	0.035	2.11	27.19	0.807	0.230	29.54	23.57	0.840	0.188	14.00
CAT-3DGS Pro	32.41	0.959	0.041	2.07	27.33	0.793	0.245	15.62	24.26	0.838	0.188	9.74
RDOGaussian	33.12	0.967	0.035	2.31	27.05	0.802	0.239	23.46	23.34	0.835	0.195	12.03
HEMGS	32.54	0.962	0.038	1.91	27.46	0.797	0.241	14.38	24.39	0.842	0.185	9.03
MesonGS-FT	32.92	0.968	0.033	3.66	26.99	0.796	0.247	27.16	23.32	0.837	0.193	16.99
HAC (lr)	33.24	0.967	0.037	1.18	27.53	0.807	0.238	15.26	24.04	0.846	0.187	8.10
HAC (hr)	33.71	0.968	0.034	1.86	27.77	0.811	0.230	21.87	24.40	0.853	0.177	11.24
ContextGS (lr)	32.79	0.965	0.040	1.01	27.62	0.808	0.237	12.68	24.20	0.852	0.184	7.05
ContextGS (hr)	33.51	0.968	0.035	1.56	27.75	0.811	0.231	18.41	24.29	0.855	0.176	11.80
CompGS (lr)	-	-	-	-	26.37	0.778	0.267	8.83	23.11	0.815	0.236	5.89
CompGS (hr)	-	-	-	-	27.26	0.803	0.239	16.50	23.70	0.837	0.208	9.60
HAC++ (lr)	33.03	0.966	0.039	0.88	27.60	0.803	0.253	8.34	24.22	0.849	0.190	5.18
HAC++ (hr)	33.76	0.969	0.033	1.84	27.82	0.811	0.231	18.48	24.32	0.854	0.178	8.63
RacerGS (lr)	32.61	0.964	0.041	0.28	27.51	0.801	0.262	3.85	24.45	0.847	0.195	4.19
RacerGS (hr)	33.78	0.970	0.035	1.44	27.84	0.811	0.243	13.61	24.50	0.855	0.183	8.42

Swipe to see all columns →

Method	DeepBlending				BungeeNeRF
Method	PSNR↑	SSIM↑	LPIPS↓	Size↓	PSNR↑	SSIM↑	LPIPS↓	Size↓
3DGS	29.42	0.899	0.247	663.9	24.87	0.841	0.205	1616
Scaffold-GS	30.21	0.906	0.254	66.00	26.62	0.865	0.241	183.0
Lee et al.	29.79	0.901	0.258	43.21	23.36	0.788	0.251	82.60
Compressed3D	29.38	0.898	0.253	25.30	24.13	0.802	0.245	55.79
EAGLES	29.72	0.906	0.249	52.34	25.89	0.865	0.197	115.2
SizeGS	29.67	0.898	0.257	6.57	26.15	0.831	0.251	19.88
LightGaussian	27.01	0.872	0.308	33.94	24.52	0.825	0.255	87.28
SOG	29.12	0.892	0.270	5.70	22.43	0.708	0.339	48.25
Navaneet et al.	29.90	0.907	0.251	13.50	24.70	0.815	0.266	33.39
Reduced3DGS	29.63	0.902	0.249	18.00	24.57	0.812	0.228	65.39
CAT-3DGS Pro	29.82	0.901	0.255	5.98	26.29	0.836	0.248	18.52
RDOGaussian	29.63	0.902	0.252	18.00	23.37	0.762	0.286	39.06
HEMGS	29.96	0.904	0.253	5.51	26.43	0.840	0.241	17.34
MesonGS-FT	29.51	0.901	0.251	24.76	23.06	0.771	0.235	63.11
HAC (lr)	29.98	0.902	0.269	4.35	26.48	0.845	0.250	18.49
HAC (hr)	30.34	0.906	0.258	6.35	27.08	0.872	0.209	29.72
ContextGS (lr)	30.11	0.907	0.265	3.45	26.90	0.866	0.222	14.00
ContextGS (hr)	30.39	0.909	0.258	6.60	27.15	0.875	0.205	21.80
CompGS (lr)	29.30	0.895	0.293	6.03	-	-	-	-
CompGS (hr)	29.69	0.901	0.269	8.77	-	-	-	-
HAC++ (lr)	30.16	0.907	0.256	2.91	26.78	0.858	0.235	11.75
HAC++ (hr)	30.34	0.911	0.254	5.28	27.17	0.879	0.196	20.82
RacerGS (lr)	30.08	0.906	0.271	1.75	26.51	0.845	0.255	7.19
RacerGS (hr)	30.36	0.913	0.256	4.48	27.18	0.867	0.216	16.15

Swipe to see all columns →

RacerGS outperforms the baselines in size reduction and fidelity preservation across diverse datasets.

BibTeX

@inproceedings{wang2026racergs,
  title={RacerGS: Residual-Autoregressive Context for 3D Gaussian Splatting Compression},
  author={Wang, Wenqing and Zeng, Huimin and Fu, Yun},
  booktitle={Advances in Neural Information Processing Systems},
  year={2026},
  eprint={2610.34263},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}