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

Wenqing Wang; Huimin Zeng; Yun Fu

cs.CVOct 25, 2026

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

Authors: Wenqing Wang, Huimin Zeng, Yun Fu

Organizations: Northeastern University

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 150x compression over vanilla 3DGS and 28x 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.

Figures & tables

Figure 2: Pipeline overview. RacerGS compresses 3DGS through three components: (1) a hash projection model maps hash encodings $\bm{f}^{h}_{i}$ to compact spatial context features $\bm{f}^{c}_{i}$ ; (2) adaptive residual quantization encodes anchor attributes relative to context-predicted means using learned quantization steps; (3) an autoregressive GRU summarizes previously decoded feature chunks, followed by FiLM modulation and a shared predictor conditioned on base statistics and chunk embeddings. The refined parameters are combined with the spatial prior to form an accurate Gaussian mixture model for residual entropy coding.

Datasets	Synthetic-NeRF				Mip-NeRF360				Tanks&Temples
Methods	psnr $\uparrow$	ssim $\uparrow$	lpips $\downarrow$	size $\downarrow$	psnr $\uparrow$	ssim $\uparrow$	lpips $\downarrow$	size $\downarrow$	psnr $\uparrow$	ssim $\uparrow$	lpips $\downarrow$	size $\downarrow$
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

Table 1: Quantitative comparisons. Storage sizes are measured in MB. The best and second-best compression results are highlighted in green and blue , respectively. Overall, RacerGS outperforms the baselines in size reduction and fidelity preservation across diverse datasets.

Figure 3: RD curve comparisons. RacerGS achieves better RD performance than the baseline methods across datasets.

Figure 4: Qualitative comparisons. RacerGS preserves details better and produces fewer artifacts than the baseline methods across scenes from diverse datasets.

$\lambda_{r}$	Total Size	$\bm{x}^{a}$	$\bm{f}^{a}$	$\bm{l}$	$\bm{o}$	$\mathcal{H}$	$\bm{m}$	MLP
$4e\text{-}3$	3.85	0.19	1.89	0.80	0.48	0.05	0.10	0.34
$0.5e\text{-}3$	13.61	0.44	8.15	2.44	1.83	0.14	0.27	0.34

Table 2: Storage sizes (MB) of different components on Mip-NeRF360 (see Appendix RacerGS: Residual-Autoregressive Context for 3D Gaussian Splatting Compression for results with additional

\lambda_{r}

values).

Table 6

Ablation Items	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size (MB) $\downarrow$	FPS $\uparrow$	Train (s) $\downarrow$
W/o residual quantization	27.83	0.810	0.242	18.24	146	2856
W/o hash projection	27.62	0.804	0.248	15.21	107	2865
Concat. context instead of GRU	27.81	0.806	0.244	16.87	156	2841
GRU context w/o FiLM	27.75	0.802	0.247	17.32	149	2854
RacerGS (full)	27.84	0.811	0.243	13.61	171	2850

Table 5: Ablation studies on the Mip-NeRF360. The best results are in bold .

Table 8

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

	PSNR (dB)	Size (MB)
None (raw $\bm{f}^{h}$ )	27.62	15.21
Linear	27.76	14.28
MLP (ours)	27.84	13.61

Appendix

Table 8: Design analyses on Mip-NeRF360. (a) Entropy-context projection. (b) Sensitivity to context dimension

D^{c}

. (c) Anchor-feature symbol statistics under attribute-lattice and residual coding. (d) Variability across three seeds (mean

\pm

std).

Dataset	Total (s)	Codec (s)	Network (s)
Synthetic-NeRF	1.75	1.08	0.67
Mip-NeRF360	14.20	8.78	5.42
BungeeNeRF	25.98	16.07	9.91

Appendix

Table 9: Decoding time breakdown . Arithmetic coding accounts for about 60% of decoding time across all datasets, with the rest spent on network evaluation.

$\lambda_{r}$	Total Time (s)	$\bm{x}^{a}$	$\bm{f}^{a}$	$\bm{l}$	$\bm{o}$	$\mathcal{H}$	$\bm{m}$	Others
$4e\text{-}3$	4.24	0.72 (17%)	1.93 (46%)	0.93 (22%)	0.45 (11%)	0.01 (0%)	0.01 (0%)	0.19 (4%)
$3e\text{-}3$	5.36	0.83 (15%)	2.59 (48%)	1.18 (22%)	0.56 (10%)	0.01 (0%)	0.01 (0%)	0.18 (3%)
$2e\text{-}3$	6.57	0.94 (14%)	3.25 (49%)	1.39 (21%)	0.69 (11%)	0.01 (0%)	0.01 (0%)	0.28 (4%)
$1e\text{-}3$	7.61	1.01 (13%)	3.88 (51%)	1.55 (20%)	0.82 (11%)	0.01 (0%)	0.01 (0%)	0.33 (4%)
$0.5e\text{-}3$	8.59	1.07 (12%)	4.40 (51%)	1.74 (20%)	0.96 (11%)	0.01 (0%)	0.01 (0%)	0.40 (5%)

Appendix

Table 10: Encoding time for different components on the Mip-NeRF360 dataset.

$\lambda_{r}$	Total Time (s)	$\bm{x}^{a}$	$\bm{f}^{a}$	$\bm{l}$	$\bm{o}$	$\mathcal{H}$	$\bm{m}$	Others
$4e\text{-}3$	7.40	0.28 (4%)	3.69 (50%)	2.30 (31%)	0.94 (13%)	0.01 (0%)	0.01 (0%)	0.17 (2%)
$3e\text{-}3$	9.15	0.31 (3%)	4.71 (51%)	2.79 (30%)	1.21 (13%)	0.01 (0%)	0.01 (0%)	0.11 (1%)
$2e\text{-}3$	10.98	0.34 (3%)	5.72 (52%)	3.21 (29%)	1.54 (14%)	0.01 (0%)	0.01 (0%)	0.15 (1%)
$1e\text{-}3$	12.63	0.37 (3%)	6.65 (53%)	3.54 (28%)	1.82 (14%)	0.01 (0%)	0.01 (0%)	0.23 (2%)
$0.5e\text{-}3$	14.20	0.40 (3%)	7.55 (53%)	3.86 (27%)	1.98 (14%)	0.01 (0%)	0.01 (0%)	0.39 (3%)

Appendix

Table 11: Decoding time for different components on the Mip-NeRF360 dataset.

$\lambda_{r}$	Total Size	$\bm{x}^{a}$	$\bm{f}^{a}$	$\bm{l}$	$\bm{o}$	$\mathcal{H}$	$\bm{m}$	MLP
$4e\text{-}3$	3.85	0.19	1.89	0.80	0.48	0.05	0.10	0.34
$3e\text{-}3$	4.92	0.20	2.57	1.01	0.63	0.06	0.11	0.34
$2e\text{-}3$	6.88	0.26	3.81	1.37	0.88	0.08	0.14	0.34
$1e\text{-}3$	10.24	0.35	5.94	1.93	1.35	0.10	0.23	0.34
$0.5e\text{-}3$	13.61	0.44	8.15	2.44	1.83	0.14	0.27	0.34

Appendix

Table 12: Storage size (MB) of different components on the Mip-NeRF360 dataset.

$\lambda_{r}$	Scenes	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size (MB) $\downarrow$
$4e\text{-}3$	chair	33.54	0.960	0.025	0.28
	drums	26.75	0.950	0.032	0.31
	ficus	34.36	0.982	0.018	0.20
	hotdog	35.65	0.971	0.033	0.24
	lego	32.76	0.966	0.029	0.36
	materials	30.21	0.950	0.025	0.31

Appendix

Table 13: RacerGS results on the Synthetic-NeRF dataset for different

\lambda_{r}

values.

$\lambda_{r}$	Scenes	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size (MB) $\downarrow$
$4e\text{-}3$	bicycle	24.19	0.743	0.292	5.42
	bonsai	32.00	0.927	0.188	2.87
	counter	29.14	0.884	0.261	2.50
	flowers	24.02	0.807	0.213	4.09
	garden	26.20	0.759	0.295	5.14
	kitchen	31.51	0.905	0.157	2.73

Appendix

Table 14: RacerGS results on the Mip-NeRF360 dataset for different

\lambda_{r}

values.

$\lambda_{r}$	Scenes	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size (MB) $\downarrow$
$4e\text{-}3$	truck	25.82	0.861	0.178	4.46
$4e\text{-}3$	train	23.08	0.833	0.212	3.92
$3e\text{-}3$	truck	25.88	0.868	0.171	5.28
$3e\text{-}3$	train	23.08	0.836	0.208	4.62
$2e\text{-}3$	truck	25.94	0.876	0.166	6.71
$2e\text{-}3$	train	23.08	0.838	0.205	5.91

Appendix

Table 15: RacerGS results on the Tanks&Temples dataset for different

\lambda_{r}

values.

$\lambda_{r}$	Scenes	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size (MB) $\downarrow$
$4e\text{-}3$	playroom	30.02	0.904	0.268	1.52
$4e\text{-}3$	drjohnson	30.13	0.907	0.274	1.97
$3e\text{-}3$	playroom	30.12	0.906	0.261	2.02
$3e\text{-}3$	drjohnson	30.28	0.910	0.265	2.68
$2e\text{-}3$	playroom	30.22	0.908	0.258	2.64
$2e\text{-}3$	drjohnson	30.32	0.910	0.260	3.52

Appendix

Table 16: RacerGS results on the DeepBlending dataset for different

\lambda_{r}

values.

$\lambda_{r}$	Scenes	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Size (MB) $\downarrow$
$4e\text{-}3$	amsterdam	26.75	0.854	0.239	8.23
	bilbao	26.32	0.878	0.251	7.32
	hollywood	26.52	0.875	0.358	8.03
	pompidou	26.73	0.822	0.255	6.27
	quebec	27.49	0.809	0.175	6.09
	rome	25.25	0.831	0.252	7.17

Appendix

Table 17: RacerGS results on the BungeeNeRF dataset for different

\lambda_{r}

values.

3DGS Compression

3D Gaussian Splatting compression reduces storage and rendering costs while preserving the quality of reconstructed views and scenes.

See all 85 papers

Towards Practical Compression of 3D Gaussian Splatting
Sep 24, 2026Pengpeng Yu, Yueru Chen, Fei Song +43DGS Compression Entropy Coding
Sun Yat-sen University, China · Pengcheng Laboratory, China · Academy of Broadcasting Science, National Radio and Television Administration, China +1
Compressing 3D Gaussian Splatting via Cross-Representation Priors
Sep 19, 2026Yezheng Zhang, Huanxiong Liang, Chuqin Zhou +23DGS Compression Gaussian Splatting Compression
Institute of Image Communication and Network Engineering, Shanghai Jiao Tong University, Shanghai 200240, China
Smaller and Faster 3DGS via Post-Training Dictionary Learning
May 28, 2026Jiarong Gong, Jonas Unger, Ehsan Miandji3DGS Compression Model Compression
Linköping University Department of Science and technology

Sep 24, 2026J/K move · Enter open · S save

Towards Practical Compression of 3D Gaussian Splatting

Pengpeng Yu, Yueru Chen, Fei Song, Tai Qin+3

Sun Yat-sen University, China · Pengcheng Laboratory, China · Academy of Broadcasting Science, National Radio and Television Administration, China+1

3DGS Compression Entropy Coding Gaussian Splatting Compression

Open paper AskPDF arXiv

3D Gaussian Splatting (3DGS) enables high-quality novel-view synthesis but requires substantial storage. Existing compression methods often rely on spatial context modeling over irregular 3D representations, increasing the complexity of training and coding. Meanwhile, floating-point context inference can introduce numerical inconsistencies across platforms, causing entropy-decoding failures. To address these practical challenges, we propose COSA-GS, which constructs context without spatial aggregation through anchor-wise causal factorization. Specifically, we use geometry context derived from each anchor's coordinates to model a compact learnable anchor latent. The anchor latent is then fused with the geometry context to form an anchor context for attribute coding. The resulting context model features a simple architecture composed solely of linear transformations and activations. We train COSA-GS using rate--distortion optimization with adaptive Gaussian pruning. Further, we develop quantization-aware training and integer inference for the context model to achieve bit-exact consistency of entropy-decoded symbols across platforms. Experiments demonstrate that COSA-GS achieves state-of-the-art compression performance while retaining fast and consistent cross-platform decoding, providing a simple yet effective framework for practical 3DGS compression. Code is available at https://github.com/pengpeng-yu/COSA-GS.

Abstract

3DGS Compression

Explore similar work

Towards Practical Compression of 3D Gaussian Splatting

Compressing 3D Gaussian Splatting via Cross-Representation Priors

Smaller and Faster 3DGS via Post-Training Dictionary Learning