Neural representation has emerged as a powerful tool to mitigate the demand for increasingly higher-resolution images. However, a fundamental trade-off remains between structured (e.g., feature grids, which offer spatial consistency) and unstructured (e.g., 2D Gaussians, which adaptively capture local variances) representations. This choice significantly affects how well an algorithm represents an image. Here, we present Gaussian-Grid, a hybrid neural representation that adaptively integrates structured and unstructured primitives. By dynamically partitioning computation between 2D Gaussian distributions and their attached multi-resolution feature grids based on the local frequency content, Gaussian-Grid automatically optimizes its architecture to match the underlying complexity of the target signal. We evaluate our approach using high-resolution natural images and demonstrate its applications as GPU-native dataset storage, showing 1.24-2.14× decoding speed improvement and PCIe data traffic reduction of 23×-176× over JPEG + ZIP data archiving. The results show that Gaussian-Grid maintains small memory footprints while achieving high reconstruction fidelity across a variety of datasets.
Figures & tables
Figure 1: Gaussian-Grid leverages Gaussian-defined feature grid layout that is guided by image frequency map to generate high quality reconstruction.
Figure 2: Overview of our method. Gaussian-Grid optimizes 2D Gaussian primitives to a spatial frequency map derived from the target image. Gaussians are then used to place feature grids of variable resolutions, aligning with the underlying frequency of the content. In the final stage, the feature grids are optimized to reconstruct the final image.
Figure 3: Pixel Rendering. For a given pixel location, Gaussian-Grid bilinearly samples top-K overlapping Gaussian grids as well as the canvas grid. The interpolated features are decoded by a small MLP dec to render the final pixel value.
Figure 4: Qualitative comparison with other recent image representation baselines. Examples are from the DIV2K dataset. Columns (a)-(e) show model reconstructions with error map insets, (f) and (g) show reference crops and the full reference images respectively. The budgets of I-NGP, GI, SAD, IMG-GS, and Gaussian-Grid are 0.1 bits per pixel respectively.
DIV2K
CLIC2020
Method
Metric
0.1 bpp
0.4 bpp
0.7 bpp
1.0 bpp
0.1 bpp
0.4 bpp
0.7 bpp
1.0 bpp
SAD [ 13 ]
PSNR ↑
24.72
29.36
31.24
32.43
26.95
31.29
33.04
34.10
SSIM ↑
0.6573
0.7816
0.8309
0.8597
0.7210
0.8096
0.8465
0.8698
MS-SSIM ↑
0.7954
0.9196
0.9495
0.9632
0.8353
0.9271
0.9518
0.9643
LPIPS ↓
0.4512
0.2980
0.2399
0.2040
0.4238
0.2996
0.2524
0.2197
GI [ 39 ]
PSNR ↑
24.21
27.27
28.71
29.69
26.36
29.15
30.40
31.33
Table 1: Comparison of Gaussian-Grid with recent primitive-based and neural baselines on DIV2K (100 images) and CLIC2020 validation (41 images) natural image datasets across four budgets (0.1 bits per pixel to 1.0 bits per pixel). The best result is in dark blue bold and the second best is in light blue.
Figure 6
Dataset size (MB)
PSNR (dB)
JPEG decode (ms / image)
Our decode (ms / image)
bpp
JPEG
Ours
JPEG
Ours
Δ
Unzip
Dec
PCIe
Total
Load
PCIe
Dq
Krnl
Total
Speedup
PCIe×
0.1
4.85
4.62
23.25
25.58
+2.33
0.87
10.05
0.73
11.65
0.17
0.14
0.64
4.48
5.44
2.14×
176×
0.4
14.71
14.85
29.05
29.36
+0.31
1.20
9.46
0.67
11.33
0.16
0.15
0.63
6.78
7.73
1.47×
55×
0.7
24.97
25.28
31.75
31.22
−0.53
1.26
9.88
0.64
11.78
0.16
0.15
0.57
8.25
9.13
1.29×
32×
1.0
35.48
35.71
33.55
32.55
−1.00
1.30
10.61
0.64
12.55
0.16
0.16
0.59
9.19
10.11
1.24×
23×
All
80.0
80.5
—
+0.27
1.16
10.00
0.67
11.83
0.16
0.15
0.61
7.18
8.10
1.46×
40×
Table 3: End-to-end decode comparison against JPEG + ZIP on 100 DIV2K images at four budgets. At 0.1 bpp our representation yields +2.33 dB over a size-matched JPEG while decoding 2.14× faster; The decode speedup ( 1.24 – 2.14× , Speedup column) persists across all budgets. The rightmost column ( PCIe× ) shows the most distinctive property: transferring quantized model weights (47–366 KB) instead of decoded pixel maps ( 8,306 KB per image) reduces CPU-to-GPU traffic by 23 – 176× . Pipeline columns. Ours: weights load ( Load ), GPU transfer ( PCIe ), GPU dequantize ( Dq ), GPU kernel forward pass ( Krnl ). JPEG: one-time ZIP extraction on CPU amortized over 100 images ( Unzip ), JPEG CPU decode ( Dec ), and pixel transfer to GPU ( PCIe ).
2D Gaussian Splatting has emerged as a novel image representation technique that can support efficient rendering on low-end devices. However, scaling to high-resolution images requires optimizing and storing millions of unstructured Gaussian primitives independently, leading to slow convergence and redundant parameters. To address this, we propose Structured Gaussian Image (SGI), a compact and efficient framework for representing high-resolution images. SGI decomposes a complex image into multi-scale local spaces defined by a set of seeds. Each seed corresponds to a spatially coherent region and, together with lightweight multi-layer perceptrons (MLPs), generates structured implicit 2D neural Gaussians. This seed-based formulation imposes structural regularity on otherwise unstructured Gaussian primitives, which facilitates entropy-based compression at the seed level to reduce the total storage. However, optimizing seed parameters directly on high-resolution images is a challenging and non-trivial task. Therefore, we designed a multi-scale fitting strategy that refines the seed representation in a coarse-to-fine manner, substantially accelerating convergence. Quantitative and qualitative evaluations demonstrate that SGI achieves up to 7.5x compression over prior non-quantized 2D Gaussian methods and 1.6x over quantized ones, while also delivering 1.6x and 6.5x faster optimization, respectively, without degrading, and often improving, image fidelity. Code is available at https://github.com/zx-pan/SGI.
Zixuan Pan, Kaiyuan Tang, Jun Xia +5
University of Notre Dame · Tohoku University · Leibniz-Institut für Analytische Wissenschaften – ISAS – e.V.
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
Recent work has shown that implicit neural representations (INRs) can be trained to effectively compress structured and unstructured volume data, allowing for direct data querying with a reduced memory footprint. However, as existing INRs for unstructured volumes do not encode geometry, they require partial mesh storage for later sampling, limiting achievable compression. At the same time, novel view synthesis methods have shown that explicit collections of 3D Gaussians can be used to accurately visualize volume data. In this work, we introduce an explicit model for volume data compression based on 3D Gaussian primitives. We reinterpret collections of 3D Gaussians as an explicit representation of a scalar field and use a sampling strategy that reconstructs scalar values at spatial locations through weighted aggregation of intersecting Gaussians. We develop optimized CUDA-accelerated pipelines for structured and unstructured model sampling, loss functions that encourage accurate domain encoding by our models, and a novel sampling-error based densification strategy. Our explicit formulation naturally encodes domain geometry, eliminating the need for mesh storage in unstructured volumes and introducing significantly higher compression opportunities. Compared to existing INRs, we demonstrate that our explicit model achieves competitive reconstruction quality with significant training speedups on structured volumes, while markedly outperforming in all metrics on unstructured volumes.