acceptodds
Withdrawn from ICLR 2027

DynaOrtho-SR: Online Orthogonal Co-Evolution for Real-World Image Super-Resolution

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Real-world image super-resolution based on pre-trained diffusion models requires balancing faithful restoration with natural detail synthesis. Spatially non-uniform degradations demand region-adaptive restoration, while continued optimization of the pixel restoration branch can cause a fixed subspace constraint to become misaligned with its evolving representation. We propose DynaOrtho-SR, a one-step diffusion framework that integrates local degradation-aware expert modeling with online orthogonal co-evolution (OOCE). A local degradation-aware convolutional mixture-of-experts (C-MoE) jointly exploits intermediate features and local degradation descriptors for spatially adaptive routing. During alternating optimization of pixel experts and semantic adapters, we periodically update the pixel restoration subspace and constrain semantic updates through gradient projection and adapter-weight retraction. Since parameter-space orthogonality does not necessarily ensure functional complementarity, we further introduce functional orthogonality regularization (F-Ortho) on shared probes to suppress correlations between branch responses, providing complementary constraints in parameter and function spaces. At inference time, adjusting branch contributions yields different trade-offs between fidelity and perceptual quality without retraining or additional iterative diffusion sampling steps. Experiments on synthetic and real-world degradation benchmarks demonstrate competitive reconstruction performance, with a favorable balance between reconstruction fidelity and reference-based perceptual quality on real-world benchmarks.

Primary area:
generative models
Keywords:
real-world image super-resolution, diffusion models, mixture of experts, degradation-aware restoration, orthogonal co-evolution

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_JbkmOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Unlike usual superresolution (SR) the proposed method claims to ake care of inhomogeneous (space varying ) blurring also. Hence they try to estimate the local departure in HR restored image and diffusion is controlled accordingly with the help of LoRA.

Critical strengths

It is worthwhile to explore how to take care of space varying blur during image SR. Use of orthogonality is also a good attempt. There is some marginal increase in the SOTA.

Critical weaknesses

I am quite disheartened at the experiments (although there are many) because the data set did not seem to have any space varying blur (at least not noticeable). Then how do claim that you have solved the problem if not experimented on the proper data. Your ablation studies should show the benefit. Currently it does not.

The orthogonality is enforced through a low rank SVD-based approximation at each step of iteration. This is huge amount of computation. Remember SVD is O(N^3) computation. I do not see any attempt to improve or computational comparison for the computation. There are classical techniques like rank-1 update etc which can speed it up.

I do not understand what is zoomed HR image for comparison. If you zoom it up, it will be a low pass one and will be inferior. Your comparison should be with the HR.

Novelty is quite low in terms of the methodology being proposed.

Questions for authors

See above.

Additionally, it will be good to see how does the SR performance degrade with increasing magnification. I do not see any experiment or ablation studies.

Reviewer_kRGBOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents DynaOrtho-SR, a one-step diffusion approach to real-world image super-resolution that aims to balance reconstruction fidelity and perceptual quality. It further leverages orthogonality constraints to decouple pixel restoration from semantic enhancement.

Critical strengths

The framework incorporates a local degradation-aware C-MoE for spatially adaptive restoration. Combining S-LoRA with C-MoE provides a flexible mechanism for handling spatially varying degradations.

Critical weaknesses

(1) The pixel-level and semantic training objectives closely resemble those in PiSA-SR. Both approaches use a pixel reconstruction loss to preserve fidelity and perceptual and distillation losses to enhance semantic restoration, making the proposed loss-based separation insufficiently distinctive. Moreover, F-Ortho provides only marginal improvements. On RealSR, it increases PSNR by just 0.0369 dB and decreases DISTS from 0.1773 to 0.1772. These negligible gains offer limited evidence that the orthogonality constraint adds meaningful value beyond training the two branches with different objectives.

(2) The quantitative results do not convincingly demonstrate an improved fidelity–perception trade-off. Table 1 presents only full-reference metrics for the fidelity-oriented setting, whereas Table 2 presents perceptual metrics for the perception-oriented setting without PSNR or SSIM. Neither table reports reconstruction fidelity alongside no-reference perceptual quality. Both settings should be evaluated against the same baselines using a consistent set of metrics, ideally with trade-off curves, to establish whether perceptual quality improves at comparable fidelity levels.

Questions for authors

Have you compared the proposed routing strategy with a content-only router that selects experts directly from input features, without using degradation descriptors? Such a comparison would help isolate the contribution of degradation-aware routing from that of MoE itself.

Reviewer_MJSuOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes DynaOrtho-SR for real world image super resolution using a single step diffusion model. It uses a degradation aware mixture of experts for spatially adaptive restoration. Online orthogonal optimization and regularization are also introduced to balance pixel fidelity and perceptual quality. Experiments on synthetic and real world benchmarks show competitive results and flexible perception distortion control.

Critical strengths

1.The paper identifies a meaningful limitation of fixed orthogonality constraints, as the pixel restoration subspace may change during training. This provides a reasonable motivation for the proposed online orthogonal optimization strategy.

2.The method achieves competitive reconstruction performance on real-world image super-resolution benchmarks. The ablation studies also show clear improvements from C-MoE and OOCE, supporting their effectiveness.

Critical weaknesses

1.There is no direct evidence showing how expert routing responds to specific local degradations such as blur, noise, and compression. The authors should provide routing visualizations or further analysis to prove that the experts truly adapt to spatial variations.

2.FreqOrtho SR already combines expert modeling with orthogonal optimization for this task. However, the paper does not compare with FreqOrtho SR. It also misses a controlled study between fixed and dynamic orthogonality constraints. This makes it hard to see the real gain of the proposed extensions.

Questions for authors

1.Can the authors show more direct evidence that CMoE routing responds to specific local degradations? Right now it is hard to tell if the experts truly adapt to spatial variations.

2.Since FreqOrtho SR already combines expert modeling with orthogonal optimization, please clarify the specific novelty of DynaOrtho SR and include a direct comparison.

3.To justify the motivation of OOCE, have the authors compared fixed orthogonality constraints against dynamically updated ones during training?

4.Looking at the third example in Figure 3, DynaOrtho SR does not seem to preserve character strokes better than PiSA SR. What is the main visual improvement here?

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

Related papers

Discussion (0)

Sign in to comment.