acceptodds
Withdrawn from ICLR 2027

Same Function, Different Geometry: Choosing Optimization Coordinates for Gaussian Representations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Optimization coordinates are part of a learning algorithm, even when they leave its represented functions unchanged. We investigate their practical value in Gaussian representations, where appearance is linear at fixed geometry but joint fitting is nonlinear. Exact appearance absorption and transported Adam/AdamW controls isolate coordinate effects from capacity. We construct energy-normalized coordinates that compensate appearance as shape changes, cancelling continuous single-atom shape–appearance coupling. With the same five-profile iRprop+ calibration allowance for peak and energy coordinates, a prospective study on 64 fresh images yields mean RGB PSNR gains of 1.242 dB over a frozen historical incumbent and 0.939 dB over independently calibrated peak coordinates, averaged across attained 2, 5 and 10-second output budgets. Both medians are positive and mean LPIPS decreases; a complete repaired-arithmetic replay retains these gains. Benefits extend to three of four domains; Aircraft and arithmetic-mean error against calibrated peak remain limitations. Timings measure warm output delivery with monitored host activity. Separately, fixed output measurements yield a sharp, separation-independent angle law and first-order minimax certificate for moving one-dimensional Gaussian clusters. Prospective selection and optimizer-payoff bounds distinguish useful geometry from useful computation. Together, these results make coordinate choice a testable design decision rather than an inherited implementation detail.

Primary area:
optimization
Keywords:
optimization, reparameterization, Gaussian representations, Gaussian splatting, preconditioning

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_XfhBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies invertible optimization coordinates for Gaussian fitting. Fixed appearance absorption is analyzed through Adam/AdamW transport. Its main practical construction compensates color as Gaussian shape changes, preserving continuous single-atom energy. Calibrated image-fitting experiments and a repaired-backward replay report gains over peak coordinates. Separate Gaussian-cluster theory and coordinate-selection studies examine coupling and computational benefit.

Critical strengths
  1. The energy study matches calibration menus and physical initialization. Its complete repaired-backward replay retains approximately 1.230 dB over the historical control and 0.926 dB over calibrated peak coordinates. Negative domains and metric tradeoffs are reported.

  2. The cluster theory goes beyond a generic preconditioning argument, identifying a worst-angle law and leading fixed-probe minimax constant. The optimizer-payoff analysis also distinguishes geometric improvements from computational gains.

Critical weaknesses
  1. The practical benefit needs clearer validation across optimization budgets and settings. Figure 3(c) shows high-gain examples at 200 updates. The early showcase should be accompanied by longer-run and equal-time comparisons on the same images. The strongest energy result remains limited to four developed domains, existing priors and 256-by-256 targets. Aircraft remains adverse, and mean RGB MSE is 1.23% higher than calibrated peak. The continuous single-atom identity also does not isolate the cause of finite-raster gains from altered update scales and numerical guards.

  2. Figure 4 has verifiable reporting problems. In panel (a), the CIFAR-100 annotation at step 200 reads 16.9192, whereas Table 6 reports 13.7312; the value near 16.92 belongs to step 400. Curves also extend and decline beyond the last reported 400-step checkpoint without explanation. These issues weaken confidence in the empirical figures and require plots generated directly from traceable numerical records.

Questions for authors
  1. How should the CIFAR-100 step-200 annotation in Figure 4(a) be reconciled with Table 6?

  2. What data, if any, do the descending curve segments beyond step 400 in Figure 4(a) represent?

  3. For the same images shown in Figure 3(c), does OA retain an advantage under longer training and matched wall-clock budgets?

  4. Can the authors provide evidence distinguishing the contribution of self-coupling cancellation from the effects of changes in update scales and numerical guards, to clarify the source of the improvement?

I would consider raising my rating if these concerns are adequately addressed.

Reviewer_ANxEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies how the parameterization of Gaussian shapes and colors affects optimization in 2D Gaussian image reconstruction. It proposes adjusting each Gaussian’s color amplitude as its size changes to preserve the squared energy. It also analyzes incorporating a fixed per-channel color scale into the optimization variables, and evaluates a method for selecting optimization coordinates. Experiments on several image datasets, including DIV2K, Flowers102, Aircraft, GTSRB, STL10, and SVHN, show that parameterization can affect reconstruction quality within a fixed optimization budget. However, the benefits of fixed color scaling are largely reproduced by optimizer tuning, while energy normalization provides additional gains that vary across datasets and metrics.

The paper proposes an interesting parameterization approach. However, the presentation is difficult to follow and obscures the contribution. The theoretical results seem to be separate from the practical method and it’s unclear how everything fits together. The usefulness of the method is also not clear from the experiments.

Critical strengths
  1. The energy-normalized parameterization has a mathematical motivation - compensating color amplitude as Gaussian size changes. Experiments show that the parameterization improves reconstruction within a fixed time budget.
  2. The paper includes useful controls and reports negative results. It shows whether optimizer tuning reproduces the benefits of fixed color scaling, reports settings where the proposed methods perform worse.
Critical weaknesses
  1. The motivation and the main contributions are unclear. The writing is difficult to follow. The paper combines fixed color scaling, energy normalization, a theory for one-dimensional Gaussian clusters, and coordinate selection without clearly explaining how these fit together.
  2. The practical benefit other than optimizer tuning is unclear. The large improvements from fixed color scaling are largely reproduced by tuning the optimizer. Energy normalization provides an additional benefit, but its strongest evaluation uses 256×256 images from four previously studied domains, with settings tuned separately for each domain. Gains are significant on GTSRB and SVHN, small on STL10, and negative on Aircraft.
Questions for authors
  1. What new insight or practical advantage does OA provide beyond choosing better learning rates?
  2. Does this method help on larger images, new datasets, or with other well-tuned optimizers?
  3. Does Theorem 1 explain or guide the method used in the experiments?
Reviewer_eedoOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper fits one image with about 2,000 2D Gaussians by per-image gradient-based optimization (Section 5). It proposes energy-normalized coordinates, in which a Gaussian's color is scaled down automatically as its shape grows (Eq. 5). With the same calibration budget, these beat the usual peak coordinates by +0.939 dB RGB PSNR on 64 images from four domains, averaged over 2, 5 and 10 s budgets (Abstract; Table 1). The paper also presents OA rescaling (Section 5.2), a theorem for 1D Gaussian clusters (Theorem 1, Section 3.2), and a per-image chart selector, SGCL (Section 5.3).

Critical strengths
  • The main comparison is calibrated fairly. Energy and peak coordinates choose from the same five-profile iRprop+ menu, and the confirmation uses 64 official-test images not used before (Section 5.1).
  • Limitations are reported openly. The paper states that Aircraft is negative and that arithmetic-mean error against calibrated peak is worse (Abstract; Section 5.1), and Section 6 lists what the guarantees and transfer tests do not cover.
Critical weaknesses

1. The setting has little practical value.

  • The paper does not explain why better optimization coordinates for 2D Gaussian image fitting matter in practice. From my understanding, Gaussian representations are mostly used in 3D. I don't see any practical values of using Gaussian sets to represent 2d images.

2. The gain is too small to matter.

  • The +0.94 dB is measured at about 44 dB PSNR (energy 44.30 vs. peak 43.36 dB; Table 1 caption), where the fitted images are already nearly identical to the target.
  • The test images are small images upsampled to 256 x 256 (Appendix R.2), which makes them easy to fit. A 512-pixel configuration loses about 0.75 dB (Section 6).
  • Most of the gain comes from GTSRB and SVHN; Aircraft is negative (Table 1). Mean squared error is 1.15% worse than with peak coordinates (Section 5.1).

3. Energy chart confusion.

I don't understand how the energy chart relates to the problem. The connection is weak and hard to justify.

4. The paper is very hard to read. The main text uses a lot of undefined terms such as "supplied covector" and "guard exposure" and jargons, and most results are in an appendix that runs to page 89.

Questions for authors
  1. What practical use does 2D Gaussian image fitting have, and why is the method not shown to help in 3D, where Gaussian representations are mainly used?
  2. Why does a +0.94 dB gain at about 44 dB matter? How much time does the energy chart save to reach a fixed quality?
  3. Beyond rescaling color and changing per-block learning rates, what is the main new contribution?
  4. Why are the gains concentrated in GTSRB and SVHN, and negative on Aircraft?
Reviewer_xdrnOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies coordinate optimization for 2D Gaussian Image representations.

Critical strengths
Critical weaknesses

The paper seems to be entirely written by an LLM and the Figures (e.g. Pipeline figure) are generated with image diffusion models and have visible artifacts.

The paper lacks a clear narrative or motivation, and its language is inconsistent and hard to understand. AI slop is the fitting term.

Questions for authors

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

Related papers

Discussion (0)

Sign in to comment.