Gaussian splatting has emerged as a competitive explicit representation for image and video reconstruction. In this work, we present P-GSVC, the first layered progressive 2D Gaussian splatting framework that provides a unified solution for scalable Gaussian representation in both images and videos. P-GSVC organizes 2D Gaussian splats into a base layer and successive enhancement layers, enabling coarse-to-fine reconstructions. To effectively optimize this layered representation, we propose a joint training strategy that simultaneously updates Gaussians across layers, aligning their optimization trajectories to ensure inter-layer compatibility and a stable progressive reconstruction. P-GSVC supports scalability in terms of both quality and resolution. Our experiments show that the joint training strategy can gain up to 1.9 dB improvement in PSNR for video and 2.6 dB improvement in PSNR for image when compared to methods that perform sequential layer-wise training. Project page: https://longanwang-cs.github.io/PGSVC-webpage/
Figures & tables
Figure 1. Progressive reconstruction of (a) Pruned Gaussian splats and (b) Layered Gaussian splats (our P-GSVC ). Both columns use the same number of splats, and the quality improves from left to right as more splats are added. Noticeable holes appear in (a), while (b) preserves scene integrity.
Figure 2. Comparison of gradient and loss over iterations on single-layer and multi-layer optimization strategies: (a) Single-layer optimization in GSVC; (b) Multi-layer trained sequentially in GSVC (the targeted layer is switched at the 50k iteration and the 100k iteration, while the previously trained layers remain frozen). One can observe multi-layer training has issues of unstable convergence and suboptimal local minima.
Figure 3. The architecture of P-GSVC , shown with three layers as an example. The base layer L0 provides the coarse representation, while the enhancement layers ΔL1 and ΔL2 progressively add additional 2D Gaussian splats. By incrementally incorporating these enhancement layers, the video quality is refined progressively, yielding higher-quality/resolution reconstructions.
Figure 4. Progressive procedure of layered Gaussians. ΔG1 and ΔG2 capture higher-frequency details and can be iteratively added to G0 , yielding progressively refined G2 .
Figure 5. Comparison of gradient and loss over iterations on randomly and cyclically joint training strategies: jointly optimizes layers (a) randomly and (b) in a cyclic order.
Figure 6. A sampled frame rendering from HoneyBee at three levels, L0 , L1 , and L2 . (a) shows quality scalability, and (b) shows resolution scalability. P-GSVC eliminates the artifacts of the Pruning method, achieving results comparable to the Monolithic method.
Dataset
Kodak
DIV-HR
No. Splats
5×103
7×103
9×103
5×104
7×104
9×104
PSNR ↑
LIG
26.6
27.3
28.1
27.8
28.7
29.5
P-GSVC
28.5
29.4
30.2
30.2
31.3
32.1
MS-SSIM ↑
LIG
0.918
0.937
0.948
0.944
0.958
0.966
P-GSVC
0.943
0.957
0.965
0.967
0.975
0.981
LPIPS ↓
LIG
0.436
0.380
0.339
0.383
0.325
0.284
Table 1. Comparison of visual quality of P-GSVC and Sequential method (LIG) on the Kodak and DIV-HR datasets with different numbers of splats.
Figure 7. Scalability evaluation on the UVG dataset, comparing (a) quality and (b) resolution across Monolithic , Sequential , and P-GSVC methods.
Figure 8. Quantized rate–distortion results of P-GSVC versus baselines on UVG dataset.
School of Science and Engineering, University of Missouri - Kansas City · Cooperative Medianet Innovation Center, Shanghai Jiaotong University · Electrical and Computer Engineering, University of Canterbury +1