acceptodds
Withdrawn from ICLR 2027

MoE-GS: π-Gated Mixture-of-Experts for Content-Adaptive 2D Gaussian Image Representation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Explicit Gaussian image representations support direct spatial rendering, but their quality–storage trade-off depends on the precision of every stored attribute. We investigate the interaction between shared appearance, quantization, and truncated rendering in MoE-GS. Each Gaussian combines a small shared color table through continuous mixture coordinates, while a learned positive amplitude participates in both top- selection and normalized blending. We formulate the storage break-even condition, characterize the color and selection constraints, and optimize the representation through its quantized inference path. A self-contained bitstream makes the reported quality reproducible from the stored bytes alone. Our evaluation distinguishes 16-bit uniform quantization from IEEE FP16, combines fixed-budget joint fitting with fixed-geometry conditional controls, and measures component interventions without changing geometry. The results delimit when appearance sharing is useful and when its quantization error outweighs its storage savings. Region queries retain the global tile partition; their cost includes candidate discovery rather than only the final rasterization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
2DGS, Image Compression, MoE

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_L6yEOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes MOE-GS, a quantization-aware shared-appearance Gaussian representation. The experiments demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The paper proposes a MoE-GS. The idea is novel

  2. The experiments demonstrate the superiority of the proposed method.

Critical weaknesses
  1. This paper is written in a hurry. The lanuguage is hard to understand.

  2. Please do not use Autoresearch.

  3. What about the trianing iteration and training time? Since this paper proposes the method of MoE for Gaussian Reconstruction. I believe the cost of training time is multiple fold.

  4. I may change my score according to other reviewers' comments.

Questions for authors
  1. Any plan to release the code and create a project page?
Reviewer_xatWOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Image-GS represents a 2D image using Gaussian primitives, each with geometry and an independent RGB color. MoE-GS extends Image-GS by representing each Gaussian’s color as a weighted combination of a small shared color set. It also assigns each Gaussian a learned amplitude that modulates both top-KK selection and its contribution to the rendered pixel. Under a common payload budget, MoE-GS uses its lower per-primitive storage cost to store more Gaussians. Empirical experiments show that the proposed MoE-GS reconstructs images more accurately than the Image-GS baseline at this budget. However, the controls do not establish that the central shared appearance mechanism proposed in the manuscript is responsible for the improvement.

Critical strengths

The empirical evaluation includes well-designed controls, and the manuscript transparently acknowledges negative findings.

Critical weaknesses
  • The manuscript does not establish a compelling advantage of Image-GS over conventional image representations or identify a downstream application that requires it. Consequently, the practical value of further compressing Image-GS with MoE-GS remains unclear.

  • MoE-GS offers limited novelty without resolving the efficiency concern. Its main components such as shared color dictionaries, soft mixture weights, and learned importance amplitudes are well-established and closely related to prior work on VQ/RVQ, sparse steered mixtures, rasterized mixtures, and SAD. Moreover, MoE-GS requires costly per-image optimization. Thus, the method provides a limited conceptual advance without demonstrating a clear computational benefit.

Questions for authors
  • What is the precise technical novelty of MoE-GS relative to the closest cited methods?
  • What practical task is Image-GS better suited for than conventional image formats, given that it is expensive to fit and must examine all Gaussians even when rendering a small image region? Accordingly, what practical benefit does MoE-GS provide?
  • For the empirical results, at a matched payload on the full dataset, does shared appearance outperform direct RGB plus amplitude?
Reviewer_3dwtOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper explained MoE-GS, a 2D Gaussian image representation targeting low-bit-rate image storage. Relative to Image-GS-style representations, it introduces two coupled modifications: (1) per-Gaussian RGB values are replaced by continuous mixture coordinates over a small shared color table, and (2) each Gaussian receives a learned positive amplitude that affects both top-K selection and normalized blending. The method is trained through a quantized inference path and stored in a self-contained bitstream, so reconstruction quality can be evaluated directly from the decoded representation.

The paper also derives basic properties of the representation, including the convex-hull restriction of the shared color table, a storage break-even condition, and the effect of the amplitude on pairwise top-K selection. The experiments emphasize careful byte accounting and distinguish uniform 16-bit quantization from IEEE FP16. At a 0.30-bpp payload target over 45 images, the proposed four-expert configuration improves mean PSNR by 0.55 dB over the authors' uniform-16 Image-GS baseline.

Critical strengths

None

Critical weaknesses
  1. The core idea has limited novelty and unclear significance for Gaussian-based image representations. The paper mainly combines a shared-appearance color parameterization with quantization-aware fitting and a learned amplitude. The experiments show modest benefits under some settings, but they do not establish a clear new representation principle or a compelling reason why this particular formulation should matter to the broader Image-GS community. Overall, the paper reads more like an empirical analysis of a specific training/decoding design than a method built around a well-motivated scientific claim.

  2. The title “MoE-GS” is misleading. The method does not appear to use a conventional mixture-of-experts architecture with expert specialization and routing.

  3. The overall writing is difficult to follow. The paper introduces many technical terms, experimental distinctions, and implementation details without first providing concise definitions or a clear high-level narrative. This makes it unnecessarily difficult to understand what problem is being solved and how the proposed components address it.

    • The abstract is confusing and does not communicate the motivation or contribution clearly. The abstract focuses heavily on implementation details and evaluation protocol before explaining why shared appearance is worth investigating in the first place.
    • The related-work section is incomplete. Development of 3D Gaussian Splatting should be mentioned, which is the motivation behind the Gaussian Image representation and subsequent works.
    • The method section is insufficiently motivated. The paper describes the shared appearance formulation and amplitude mechanism, but does not convincingly explain how the authors arrived at these design choices, what specific limitation of existing Image-GS representations they are intended to solve, or why these choices should be expected to improve rate-distortion behavior.
    • The training and decoding pipeline in Figure 1 is hard to understand. The arrows and transitions are not sufficiently explained. The figure contains mostly generic phrases such as “quantized fields,” “independent decoder,” and “reconstruction.” .
  4. Several variables are not well defined.

    • What is pp at line 105?
    • What are WW and HH at line 102?
    • What is zz at line 138?
Questions for authors
  • What’s the motivation for using a shared appearance?
  • What is the importance of the experiment's results and analysis? How does it contribute to the area of Image-GS representation?

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

Related papers

Discussion (0)

Sign in to comment.