2D Gaussian representations provide an explicit and efficient way to represent images, but their rate-distortion performance under limited budgets depends critically on how Gaussian capacity and parameter precision are allocated. Existing methods typically optimize these two resources separately, overlooking content-dependent capacity demands and heterogeneous quantization sensitivity. We present CP-GS, a capacity-precision coupled Gaussian splatting framework for budget-aware 2D Gaussian image representation that jointly optimizes capacity allocation and attribute precision under a shared rate constraint. CP-GS first constructs a content-adaptive Gaussian representation by dynamically reallocating a fixed Gaussian budget from redundant regions to areas with higher representation demand. It then introduces hierarchical distortion-aware representation quantization, using efficient rendering-based calibration to allocate precision across Gaussian attributes and further refine it within heterogeneous attribute groups. Finally, it jointly optimizes Gaussian capacity and attribute precision by constructing candidate representations with different Gaussian counts and precision configurations and selecting operating points according to decoded distortion and actual serialized size. Experiments on DIV2K and Kodak show that CP-GS achieves improved rate-distortion performance over representative 2D Gaussian and implicit neural representations across a range of budgets, while maintaining a fixed overall representation budget and low optimization cost. These results demonstrate the benefit of jointly optimizing spatial capacity and parameter precision for compact Gaussian image representation.
Figures & tables
Figure 1: CP-GS at a glance. (a) Existing methods optimize Gaussian capacity and attribute precision separately. (b) We identify capacity mismatch, heterogeneous precision sensitivity, and capacity-precision coupling under limited budgets. (c) CP-GS performs content-adaptive capacity reallocation, hierarchical distortion-aware quantization, and joint capacity-precision optimization.
Figure 2: Illustration of our CP-GS framework. Given an input image, an initial Gaussian representation is first obtained via Image-GS. Two parallel branches, content-adaptive 2D Gaussian representation and hierarchical distortion-aware representation quantization, are then jointly optimized under a shared budget, yielding a capacity–precision coupled Gaussian representation that supports efficient decoding and downstream tasks.
Methods
Kodak
DIV2K
PSNR
MS-SSIM
Params(M)
Optim. Time (s)
PSNR
MS-SSIM
Params(M)
Optim. Time (s)
INR-based
SIREN Sitzmann et al. (2020)
26.50
0.8705
3.74
889
30.01
0.9630
0.42
2242
SIREN+ Tsym Zhang et al. (2025b)
32.80
0.9820
0.10
166
29.61
0.9712
0.10
261
EVOS Zhang et al. (2025c)
33.64
0.9809
0.10
106
30.38
0.9687
0.10
189
GS-based
Number of Gaussians (5K)
Number of Gaussians (10K)
Table 1: Comparison with representative INR- and GS-based image representation methods on Kodak and DIV2K. GS-based methods are evaluated at 5K/10K Gaussians on Kodak and 10K/30K on DIV2K. We report PSNR, MS-SSIM, parameter count, and optimization time. Optimization time denotes the single-image wall-clock time required to obtain the final representation.
CP-GS Components
BD-Rate ↓ vs. Spatial (%)
Quality Metrics
Spatial Realloc.
Attr. RD
Hier. Color
Joint Capa.–Prec.
PSNR ↑
SSIM ↑
LPIPS ↓
✓
–
19.310
0.6457
0.3699
✓
✓
−14.21
22.740 ( +3.430 )
0.6951 ( +0.0494 )
0.3639 ( −0.0060 )
✓
✓
✓
−23.71
23.225 ( +3.915 )
0.7053 ( +0.0596 )
0.3371 ( −0.0328 )
✓
✓
✓
✓
−44.93
23.781 ( +4.471 )
0.7147 ( +0.0690 )
0.3308 ( −0.0392 )
Table 2: Cumulative ablation of CP-GS on Kodak24. Quality metrics are averaged over all 24 images, with changes relative to the Spatial baseline shown in parentheses. BD-rate is measured relative to Spatial.
Figure 3: Visualization of content-adaptive Gaussian capacity reallocation on four representative images (a)–(d). Rows show the original image, Gaussian reallocation, and the pre-reallocation demand map. Red crosses denote merged Gaussians, and blue circles denote newly reallocated Gaussians.
(a) Attribute Precision Allocation
(c) Capacity-Precision Coupling
Method
Eval.
DIV2K (%)
Kodak (%)
Method
DIV2K (%)
Kodak (%)
Uniform (baseline)
0
0.00
0.00
Baseline
0.00
0.00
Fixed mixed
0
+21.35
+20.65
Hier. Precision
-7.18
-7.59
Greedy
–
-5.84
-13.11
Count-only
-32.91
-48.12
Random
16
-8.47
-3.19
Precision-only
-17.00
-14.30
Oracle
216
-17.94
-14.33
Joint RD
-45.01
-50.14
Table 3: Ablation analysis of precision allocation and capacity-precision coupling. Physical BD-rate (%, lower is better) is measured relative to Uniform in (a), Global in (b), and the fixed-count uniform baseline in (c). In (d), Structured-19 is compared with Oracle-216 under matched nominal payload targets; PSNR and BD-rate gaps are reported relative to Oracle-216.
Figure 4: Image-level BD-rate distributions of Precision-only, Joint RD, and Joint Hierarchical RD on (a) DIV2K and (b) Kodak24. Each point denotes one image, and the boxes summarize the corresponding distributions. A more negative BD-rate indicates a better rate-distortion efficiency.
2D Gaussian representations provide an explicit and lightweight alternative for image representation, enabling fast deterministic decoding without a heavy neural synthesis network. However, existing per-image Gaussian codecs still require costly image-specific optimization to determine how limited Gaussian capacity should be distributed across heterogeneous regions. We introduce CaliSplat, an optimization-efficient Gaussian image codec that selectively explores candidate representations and globally allocates a constrained Gaussian budget. We exploit the closed-form reconstruction error of the single-Gaussian case as an inexpensive measure of local difficulty, fitting higher-capacity candidates only in regions where additional Gaussians are likely to be useful. To reduce selection errors caused by mismatched convergence rates across candidate sizes, a small set of longer-horizon fits is used to calibrate short-horizon distortion estimates before global budget-aware selection. The selected representation is then compressed into a bitstream. Across standard and large image datasets, CaliSplat achieves competitive or superior rate--distortion performance while reducing per-image fitting time by 3.0×--15.7× compared with representative Gaussian codecs. It further maintains high-throughput deterministic decoding at 175.9--478.2~MP/s. These results demonstrate that content-adaptive Gaussian compression can substantially reduce per-image optimization cost while preserving strong reconstruction quality and efficient explicit decoding.
Zhenbang Zhang, Zhengquan Luo, Kai Zhao +1
Mohamed bin Zayed University of Artificial Intelligence
2D Gaussian Splatting is an attractive direction for image representation due to its explicit formulation, fast rasterization, and favorable decoding efficiency. The representation quality of this paradigm depends on the proper allocation of Gaussian capacity to the demanding regions. However, existing methods fail to allocate Gaussian capacity efficiently during optimization: under-reconstructed content is often refined in a fragmented pixel-wise manner, while neighboring optimized Gaussians with similar attributes are redundantly retained. This inefficiency motivates the need for a density control framework that jointly addresses insufficient allocation in under-reconstructed regions and redundant allocation in over-reconstructed regions. Our key insight is that this framework should exploit two complementary forms of locality: the local continuity of reconstruction errors in image space for improved Gaussian allocation, and the local similarity of neighboring Gaussians in Gaussian space for redundant elimination. Based on this insight, we propose Locality-Aware Density Control (LocoADC), a plug-and-play framework that improves Gaussian capacity utilization through Region-wise Gaussian Densification (RGD) and Similarity-Driven Gaussian Merging (SDGM) strategies, together with a local color consistency constraint for more reliable merging. Extensive experiments on diverse datasets show that LocoADC consistently improves multiple baselines by enabling more effective local Gaussian allocation, including a 2.93 dB PSNR gain over GI on the CLIC dataset under the same 30k Gaussian budget. Code is available at: https://github.com/ChenJiaCong-1005/LocoADC.
Jiacong Chen, Qingyu Mao, Xiandong Meng +5
Shenzhen University Shenzhen, China · Pengcheng Laboratory Shenzhen, China · Harbin Institute of Technology (Shenzhen) Shenzhen, China
Gaussian-based image representations effectively model image content using compact parametric primitives while preserving high visual fidelity, yet storing a large number of floating-point parameters per primitive degrades rate-distortion efficiency at higher fidelity targets. To improve the rate-distortion performance in Gaussian representation, we present our Cluster-Guided Vector Quantization (CGVQ), a Gaussian primitive based image compression method. Our key idea is to partition Gaussian parameters further into homogeneous groups prior to quantization, enabling higher compression efficiency and accurate parameter reconstruction. In practice, our extensive experiments show that CGVQ decreases the bpp by 20% with respect to our baseline, while maintaining on-par visual quality
Runze Cheng, Yicheng Zhan, Josef Spjut +1
University College London · London, UK · NVIDIA +1