acceptodds
Under review as a conference paper at ICLR 2027

Marrying Capacity and Precision for Efficient 2D Gaussian Image Representation

Abstract

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.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
2D Gaussian Representation, Image Representation, Adaptive Quantization, Budget-Aware Optimization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_cyAqOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CP-GS which is a capacity-precision coupled framework for efficient 2D Gaussian image representation. The method combines content-adaptive Gaussian reallocation, hierarchical distortion-aware quantization, and joint capacity-precision optimization to improve rate-distortion performance under storage constraints. Experiments on Kodak and DIV2K demonstrate improved reconstruction quality and compression efficiency over several existing approaches.

Critical strengths

The paper addresses a meaningful problem by jointly considering Gaussian representation capacity and attribute precision under limited storage budgets.

The proposed method achieves competitive reconstruction quality and rate-distortion performance compared with existing Gaussian-based image representation methods, while maintaining relatively low optimization costs.

The paper provides comprehensive ablation studies demonstrating the effectiveness of individual components. In particular, the structured precision allocation achieves performance close to exhaustive search with substantially fewer evaluations.

Critical weaknesses
  1. Insufficient comparison under matched compression budgets. The paper emphasizes rate-distortion optimization under storage constraints. However, the main comparison in Table 1 primarily evaluates reconstruction quality under matched Gaussian counts and parameter counts, rather than actual compressed bitstream sizes. Although the authors provide rate-distortion ablations, direct comparisons with existing Gaussian-based compression methods under matched bitrates are limited. Such comparisons would better establish the practical compression advantages of CP-GS.

  2. Limited justification of the proposed capacity-precision coupling. The joint optimization primarily constructs a finite candidate space with different Gaussian counts and quantization configurations, followed by rate-distortion-based selection. While the ablations demonstrate improvements over capacity-only and precision-only optimization, it remains unclear whether the proposed coupling significantly outperforms a simpler sequential optimization strategy that independently optimizes capacity and precision. A direct comparison would help clarify the methodological contribution of the proposed joint optimization.

  3. Insufficient clarification of the optimization cost. The proposed method starts from an Image-GS representation and introduces additional capacity reallocation, recovery optimization, precision calibration, and candidate evaluation. However, it is not sufficiently clear how the cost of these stages is accounted for in the reported optimization times. A detailed runtime breakdown, including the initial Image-GS fitting process, would help establish the fairness of the efficiency comparisons.

Questions for authors

Can the authors provide comparisons with existing Gaussian-based compression methods under matched actual bitrates? How does the proposed joint optimization compare with a sequential capacity-then-precision optimization strategy? Does the reported optimization time include the initial Image-GS fitting stage?

1 comment

The authors’ response addresses most of my concerns. After considering it, I have decided to keep my original score.

Reviewer_V86yOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces further optimizations to ImageGS: on the one hand, it reallocates the Gaussian distribution in ImageGS; on the other hand, it optimizes the quantization bit width for the parameters of each Gaussian. By setting a more appropriate bit width, it achieves higher-quality compression.

Critical strengths
  1. A reallocation method has been proposed to optimize the Gaussian distribution in ImageGS, reducing redundancy in certain areas and redistributing these Gaussians to areas where more detail needs to be represented.
  2. An adaptive parameter quantization method has been proposed, which adaptively assigns different bit widths to different parameters in order to achieve a higher compression ratio.
Critical weaknesses
  1. As the first major contribution of this paper, this reallocation does not appear to yield a significant improvement; in the images, the difference is virtually imperceptible to eyes. The improvement achieved by reallocation is too small compared to the random approach—only 0.3 dB—a difference that is difficult to discern.
  2. This is not to say that this work is meaningless; it has indeed further improved the compression performance of Gaussian-based images. However, this is not the primary bottleneck preventing such methods from being widely adopted for image compression. I still believe that for ImageGS-like methods to replace existing formats such as JPEG, it must offer some significant advantages or demonstrate a clear trend towards adoption. Its biggest bottleneck at present is its slow compression speed; no one is willing to wait several minutes to compress a single image, especially given that this is lossy compression. Alternatively, it should be employed in more suitable applications, such as texturing in 3D reconstruction, where discrete primitives are used to represent fragmented, discrete texture maps.
  3. The experiment described in this paper is rather odd and seems somewhat unfair. The authors used ImageGS as input, but perhaps its processing time wasn’t taken into account—otherwise, how could it have been faster than ImageGS itself? Furthermore, if the aim of Table 2 is truly to provide a complete cumulative ablation, then it would indeed be more reasonable for the first row to list the ImageGS baseline, rather than the version with Spatial Reallocation.
Questions for authors

Apart from the main issues mentioned in weaknesses, there are also a few other minor issues:

  1. The authors have performed redistribution based on the imageGS distribution; how does this differ from the deletion and splitting operations in 3DGS itself, and how to ensure the stability of its training?
  2. During reallocation, the author calculated a residual direction; however, could this residual direction not only represent the landing point for a new Gaussian distribution but also the direction for optimizing the existing Gaussian? Why can’t it optimize itself in that direction? On the other hand, I believe that if the residual direction is rather chaotic, this actually indicates that a new Gaussian is required—that a single Gaussian cannot handle this situation. I do not understand why you think that a chaotic residual direction actually suggests a lower need for a new Gaussian. For example, suppose the target image is a black-and-white checkboard, while the current reconstruction consists entirely of medium grey: if the residual is calculated in this way, the requirement would be very low; however, it is clearly impossible to represent these two distinctly different colors using a single Gaussian.
  3. Figure 2 isn’t very well drawn; it doesn’t illustrate the relationship between the initial Gaussian and those shown in the content-adaptive 2D Gaussian representation.
Reviewer_yQ83OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a 2D Gaussian image representation technique by jointly considering the Gaussain capacity and the precisions. It first constructs the Gaussian representation by dynamically reallocating the budget to areas with higher representation demand. It then introduces hierarchical distortion-aware representation quantization to allocate precision across Gaussian attributes. Finally, it jointly optimizes Gaussian capacity and attribute precision by constructing candidate representations and selecting operating points according to decoded distortion and actual serialized size.

Critical strengths

Motivation is convincing.

Critical weaknesses
  1. The novelty claim regarding the joint treatment of Gaussian capacity and attribute precision requires substantial clarification, since there are already several works targeting at the same goal.
  2. The clarity regading the proposed technique needs lots of improvement, especially for Section 4.3.
  3. The joint distortion is approximated by [5] without justification.
  4. For my understanding, the joint capacity and precision optimization is kind of full search since the primitives are selected first followed by attribute precision assignments. The novelty is limited.
  5. For experimental results, Table I considers only the Gaussian counts and no evaluation on attribute precision.
Questions for authors

More comparison to similar works. Table 2 should provide more information about the configuration to reveal the benefit from joint optimization.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.