3D Gaussian Splatting (3DGS) enables high-quality novel view synthesis but incurs high storage and transmission costs due to dense Gaussian primitives. Recent anchor-based compression reduces per-primitive redundancy, yet redundancy across anchors remains largely unexploited. We propose CRP-GS (Cross-Representation Priors for Gaussian Splatting), a rate-distortion optimized compression framework that leverages cross-representation priors to improve anchor-level entropy modeling. First, a Correspondence-Oriented Hierarchical Structure (COHS) organizes anchors by feature correspondence rather than spatial proximity, constructing root-leaf dependencies so that selected anchors can act as informative priors to conditionally encode others, yielding more accurate likelihood prediction and lower conditional entropy. Second, Shared Feature Aggregation (SFA) extracts globally shared features from a contextual hash grid and injects them into anchor representations, factoring out scene-consistent low-frequency information that would otherwise be redundantly embedded in individual anchors. Both modules are trained under a unified rate-distortion objective to balance bitrate reduction and rendering fidelity. Experiments across multiple benchmarks show that CRP-GS achieves a favorable overall rate-distortion trade-off, yielding around 30% average bitrate reduction compared to anchor-based baselines while maintaining comparable rendering quality.
Figures & tables
Fig. 2: CRP-GS employs cross-representation priors to jointly optimize rendering quality and bitrate. The rate path organizes anchors into a Correspondence-Oriented Hierarchical Structure (COHS) for conditional entropy coding, while the distortion path constructs enriched anchors via Shared Feature Aggregation (SFA) for high-fidelity rendering. Both paths are optimized under a unified rate-distortion training scheme.
Fig. 3: CDF plot of the number of correspondence anchors under different distances and thresholds. Distances are normalized by the scene diagonal length, defined as the diagonal of the anchor-set axis-aligned bounding box. The curves are computed from 10,000 uniformly sampled query anchors from the Bicycle scene of Mip-NeRF360 [ 57 ] , and the legend values denote cosine-similarity thresholds.
Fig. 4: Effect of SFA. Rendering with only the individual anchor feature ( fi , left) exhibits missing structures and local inconsistencies. The shared feature queried from the hash grid ( fs , middle-left) captures coarse but globally consistent scene information. Aggregating fs with fi yields the aggregated feature ( fa , middle-right), producing a more coherent reconstruction that better matches the ground truth (right).
Methods
Mip-NeRF360
Tanks&Temples
DeepBlending
BungeeNeRF
PSNR ↑
SSIM ↑
LPIPS ↓
SIZE ↓
PSNR ↑
SSIM ↑
LPIPS ↓
SIZE ↓
PSNR ↑
SSIM ↑
LPIPS ↓
SIZE ↓
PSNR ↑
SSIM ↑
LPIPS ↓
SIZE ↓
3DGS
27.49
0.813
0.222
744.7
23.69
0.844
0.178
431.0
29.42
0.899
0.247
663.9
24.87
0.841
0.205
1616
Scaffold-GS
27.50
0.806
0.252
253.9
23.96
0.853
0.177
86.50
30.21
0.906
0.254
66.00
26.62
0.865
0.241
183.0
EAGLES
27.15
0.808
0.238
68.89
23.41
0.840
0.200
34.00
29.72
0.906
0.249
52.34
25.89
0.865
0.197
115.2
LightGaussian
27.00
0.799
0.249
44.54
22.83
0.822
0.220
22.43
27.01
0.872
0.308
33.94
24.52
0.825
0.255
87.28
Compressed3D
26.98
0.801
0.238
28.80
23.32
0.832
0.194
17.28
29.38
0.898
0.253
25.30
24.13
0.802
0.245
55.79
TABLE I: 3DGS Compression Benchmark Comparison
Datasets
vs. HAC
vs. ContextGS
vs. HAC++
vs. HEMGS
vs. PCGS
BD-rate (%) ↓
BD-PSNR (dB) ↑
BD-rate (%) ↓
BD-PSNR (dB) ↑
BD-rate (%) ↓
BD-PSNR (dB) ↑
BD-rate (%) ↓
BD-PSNR (dB) ↑
BD-rate (%) ↓
BD-PSNR (dB) ↑
Mip-NeRF360
-39.28
+0.334
–
+0.204
-1.97
+0.047
-8.99
+0.050
-14.32
+0.083
Tanks&Temples
-56.90
+0.586
-52.41
+0.420
-24.31
+0.208
-6.82
+0.035
-36.02
+0.277
DeepBlending
-44.25
+0.591
-30.78
+0.208
-2.30
+0.045
-8.02
+0.054
-23.76
+0.171
BungeeNeRF
-27.06
+0.595
+15.24
-0.147
+32.56
-0.282
–
–
+41.96
-0.287
Average
-41.87
+0.526
-22.65
+0.171
+0.99
+0.005
-7.94
+0.046
-8.04
+0.061
TABLE II: BD-rate and BD-PSNR comparison over the overlapping rate-distortion range.
Fig. 5: The rate-distortion curves comparing our CRP-GS with baseline methods. Note that file size directly reflects the entropy-coded bitstream and is proportional to the effective bitrate under identical rendering settings.
Fig. 6: Qualitative comparisons of “counter” from Mip-NeRF360 [ 57 ] and “truck” from Tanks&Temples [ 60 ] .
Method
Train [ 60 ]
Room [ 57 ]
DrJohnson [ 59 ]
PSNR ↑
Model Size ↓
PSNR ↑
Model Size ↓
PSNR ↑
Model Size ↓
Our full model
22.87
4.67
32.01
3.43
29.72
3.71
Our w/o SFA
22.84 (-0.03 dB)
5.73 (+22.7%)
31.91 (-0.10 dB)
3.80 (+11.0%)
29.70 (-0.02 dB)
3.91 (+5.4%)
Our w/o COHS and SFA
22.55 (-0.32 dB)
5.74 (+22.9%)
31.64 (-0.37 dB)
3.94 (+15.0%)
29.60 (-0.11 dB)
3.93 (+6.0%)
Our w/o COHS, SFA, and AC
22.55 (-0.32 dB)
6.59 (+41.1%)
31.64 (-0.37 dB)
4.70 (+37.2%)
29.60 (-0.11 dB)
4.51 (+21.7%)
TABLE III: Ablation study on representative scenes from different datasets.
Method
PSNR (dB)
Model Size (MB)
Full Method (Ours)
22.41
5.89
Random Link
22.35
6.41
Distance Link
22.28
6.22
No Link
22.30
6.13
TABLE IV: Comparison of COHS with alternative linking strategies
Fig. 7: (a) Ablation study of the SFA mechanism; (b) Compression heatmap analysis with and without SFA.
Fig. 8: Bitstream breakdown comparison between CRP-GS and HAC.
Item
Size (MB)
Ratio (%)
Active-anchor Mask
0.03
0.60
Link Vector L
0.16
3.14
Total COHS Side Information
0.19
3.74
TABLE V: COHS side information overhead in the final bitstream
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.
Pengpeng Yu, Yueru Chen, Fei Song +4
Sun Yat-sen University, China · Pengcheng Laboratory, China · Academy of Broadcasting Science, National Radio and Television Administration, China +1
3D Gaussian Splatting (3DGS) enables high-quality, real-time novel view synthesis but requires significant storage and transmission resources. Mapping Gaussian attributes to 2D representations enables efficient compression using mature image codecs. However, Gaussian representations that achieve comparable rendering quality are generally treated similarly during optimization, despite substantial differences in coding rate, leading to limited rate-distortion (R-D) performance after compression. To address this limitation, we propose RDCompGS, a 3DGS compression method that enables overall R-D optimization of Gaussian representations with respect to subsequent coding. A differentiable estimate of the coding rate is incorporated into the R-D optimization with rendering distortion. The rate constraint adjustment model further regulates the R-D trade-off across different compression conditions. Extensive experiments on three benchmark datasets demonstrate that RDCompGS achieves average Bjøntegaard Delta Rate reductions of 37.84% and 49.31% with H.265/HEVC and JPEG XL, respectively. The estimated rate also shows a strong correlation with the actual bit-rate under different compression conditions.
Huiping Ren, Shuai Wan, Luqian Bai +2
School of Electronics and Information, Northwestern Polytechnical University, Xi’an 710129, China · Department of Computer Science, City University of Hong Kong, Hong Kong SAR, China · School of Telecommunication Engineering, Xidian University, Xi’an 710071, China
Although 3D Gaussian Splatting (3DGS) enables high-fidelity real-time rendering, its prohibitive storage overhead severely hinders practical deployment. Recent anchor-based 3DGS compression schemes reduce gaussian redundancy through some advanced context models. However, they overlook explicit geometric dependencies, leading to structural degradation and suboptimal ratedistortion performance. In this paper, we propose a Local Geometry-aware Hierarchical Context Compression framework for 3DGS(LG-HCC) that incorporates inter-anchor geometric correlations into anchor pruning and entropy coding for compact representation. Specifically, we introduce an Neighborhood-Aware Anchor Pruning (NAAP) strategy, which evaluates anchor importance via weighted neighborhood feature aggregation and then merges low-contribution anchors into salient neighbors, yielding a compact yet geometry-consistent anchor set. Moreover, we further develop a hierarchical entropy coding scheme, in which coarse-to-fine priors are exploited through a lightweight Geometry-Guided Convolution(GG-Conv) operator to enable spatially adaptive context modeling and rate-distortion optimization. Extensive experiments show that LG-HCC effectively alleviates structural preservation issues,achieving superior geometric integrity and rendering fidelity while reducing storage by up to 30.85x compared to the Scaffold-GS baseline on the Mip-NeRF360 dataset
Xuan Deng, Xiandong Meng, Hengyu Man +4
Harbin Institute of Technology, Peng Cheng Laboratory Shenzhen, China · Peng Cheng Laboratory Shenzhen, China · Harbin Institute of Technology Harbin, China +1