FreqOrtho-SR: Frequency-Guided Orthogonal Expert Learning for Real-World Image Super-Resolution

Daeyoung Kim; Dam Hoang Phuong; Dinh Phu Tran; Minh Son Hoang; Quyen Nguyen Duc

REVIEW 1 major objections 2 minor 14 references

FreqOrtho-SR: Frequency-Guided Orthogonal Expert Learning for Real-World Image Super-Resolution

T0 review · 1 major / 2 minor · reviewed 2026-06-30 · grok-4.3

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read FreqOrtho-SR separates pixel fidelity from semantic quality by routing inputs to frequency-specialized experts and projecting semantic gradients into the null space of a pixel subspace extracted via SVD.

desk verdict FreqOrtho-SR adds FFT-based non-parametric routing to LoRA experts and an SVD null-space projection for dual-objective separation in real-world SR, but the projection's claimed guarantee of non-interference rests on an alignment between weight deltas and pixel objectives that is not automatic. read the letter →

arxiv 2606.28745 v1 pith:QJVPDDPN submitted 2026-06-27 cs.CV

Minh Son Hoang , Dinh Phu Tran , Quyen Nguyen Duc , Dam Hoang Phuong , Daeyoung Kim This is my paper

classification cs.CV

keywords real-worldimagesuper-resolutionmixtureofLoRAexpertsorthogonalgradientprojectionfrequencydomainroutingSVDsubspacepixel-semantictrade-offdegradationadaptation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper seeks to resolve interference in dual-objective training for real-world image super-resolution, where semantic branches can still encroach on pixel subspaces even when modules are frozen, and where single adapters fail on varied degradations. It introduces frequency-guided mixture of LoRA experts that extract degradation signatures through FFT to route inputs to specialized experts. The orthogonal gradient projection step extracts a pixel-fidelity subspace from SVD on combined expert weight deltas and projects semantic gradients onto the orthogonal complement. This construction is claimed to produce genuinely non-interfering updates. A sympathetic reader would expect the result to be a clearer fidelity-perception trade-off and more stable adaptation across real corruptions than prior dual-branch methods.

What carries the argument

Orthogonal Gradient Projection (OGP), which extracts the pixel-fidelity subspace via SVD on combined expert weight deltas and projects semantic gradients onto its null space to enforce orthogonality.

What would settle it

A direct comparison showing whether semantic training after the SVD-based projection still produces the same shifts in pixel-level reconstruction error as non-projected training on the same expert set.

Watch

Extended reading notes

Core claim

FreqOrtho-SR comprises FreqMoE, a non-parametric FFT-based degradation-feature extractor that routes inputs to specialized LoRA experts according to frequency-domain signatures, and OGP, which reframes dual-objective optimization as a subspace-constrained problem by extracting the pixel-fidelity subspace via SVD on combined expert weight deltas and projecting semantic gradients onto its null space, thereby guaranteeing orthogonality between the two objectives.

Load-bearing premise

The SVD performed on the combined expert weight deltas will isolate a pixel-fidelity subspace whose null space remains useful and non-interfering for semantic learning.

Editorial extensions

If this is right

Achieves competitive overall performance and a strong fidelity-perception trade-off across multiple benchmarks.
Enables stable and interpretable specialization across different corruption types via frequency signatures.
Supports efficient single-step inference.
Produces genuinely complementary learning without mutual interference between objectives.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same SVD-null-space construction could be tested on other multi-objective fine-tuning settings where one loss must not alter a preserved capability.
Frequency routing based on FFT signatures might extend to tasks that classify or adapt to domain shifts defined by spectral properties.
If the extracted subspace proves stable across different numbers of experts, the method could be applied with minimal retuning to new degradation distributions.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

read the letter

The paper's concrete contribution is the pairing of a frequency-domain router that sends inputs to specialized LoRA experts without training a router, plus the OGP step that runs SVD on the concatenated expert deltas and projects semantic gradients into the orthogonal complement.

The FFT routing looks useful on its own. It uses a fixed, interpretable feature to handle diverse real-world degradations instead of a learned gate, which avoids extra parameters and gives stable specialization. The code release helps here.

The OGP construction is the part that needs scrutiny. The abstract states that SVD on the combined deltas isolates a pixel-fidelity subspace whose null space then keeps semantic updates from interfering. That would be a clean way to avoid the capacity-expansion problem in frozen dual-branch setups. However, the principal directions of the LoRA deltas reflect the low-rank update geometry, not necessarily the directions that most affect pixel loss. If those top singular vectors do not line up with actual pixel gradients, the projected semantic steps can still degrade fidelity while staying formally orthogonal in parameter space. The abstract supplies no derivation or check that the subspace matches the pixel objective.

No quantitative results, ablations, or dataset details appear in the provided text, so the competitive performance claim cannot be assessed yet. The method is algorithmic rather than self-referential, and the authors engage directly with the dual-branch interference issue in the literature.

This is for readers working on real-world SR or parameter-efficient multi-objective training. It is worth sending to peer review so the alignment assumption and the empirical trade-off can be tested.

Referee Report

1 major / 2 minor

Summary. The manuscript proposes FreqOrtho-SR for real-world image super-resolution. It introduces FreqMoE, a frequency-guided mixture of LoRA experts that routes inputs via a non-parametric FFT-based degradation-feature extractor to enable specialization across corruption types, and OGP, which extracts the pixel-fidelity subspace via SVD on combined expert weight deltas and projects semantic gradients onto its null space to guarantee orthogonality between pixel-fidelity and semantic objectives. The approach claims to achieve competitive performance and a strong fidelity-perception trade-off across benchmarks with efficient single-step inference; source code is provided.

Significance. If the OGP construction reliably isolates a pixel-fidelity subspace aligned with the loss landscape and the null-space projection enforces non-interference, the method would offer a principled algorithmic solution to multi-objective optimization in diffusion-based SR, addressing a persistent challenge in balancing fidelity and perception. The frequency-domain routing for expert specialization could also improve adaptation to heterogeneous real-world degradations. The public release of source code supports reproducibility.

major comments (1)

[OGP description (abstract)] OGP description (abstract): the claim that SVD on the matrix of concatenated LoRA deltas 'guarantees orthogonality between the two objectives, enabling genuinely complementary learning without mutual interference' rests on the unverified assumption that the top right singular vectors span the directions most affecting pixel-level fidelity. These vectors instead reflect the geometry of the low-rank LoRA parameterization; if they are misaligned with the pixel-loss landscape (or also affect semantic features), the projected semantic gradients can still degrade pixel metrics while remaining formally orthogonal in parameter space. No derivation, alignment check, or ablation is supplied to establish that the extracted subspace matches the pixel objective.

minor comments (2)

[Abstract] Abstract: asserts 'competitive overall performance and a strong fidelity-perception trade-off' but supplies no quantitative results, ablation tables, error bars, dataset details, or baseline comparisons, preventing verification of the central empirical claims.
[Abstract] Abstract: the phrasing 'non-parametric FFT-based degradation-feature extractor' is introduced without a concrete description of the feature computation or routing rule, leaving the specialization mechanism underspecified.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive comment regarding the justification of the OGP subspace construction. We address the concern directly below.

read point-by-point responses

Referee: the claim that SVD on the matrix of concatenated LoRA deltas 'guarantees orthogonality between the two objectives, enabling genuinely complementary learning without mutual interference' rests on the unverified assumption that the top right singular vectors span the directions most affecting pixel-level fidelity. These vectors instead reflect the geometry of the low-rank LoRA parameterization; if they are misaligned with the pixel-loss landscape (or also affect semantic features), the projected semantic gradients can still degrade pixel metrics while remaining formally orthogonal in parameter space. No derivation, alignment check, or ablation is supplied to establish that the extracted subspace matches the pixel objective.

Authors: We appreciate the referee's precise observation. The OGP procedure constructs the subspace exclusively from the weight deltas of experts trained under the pixel-fidelity loss; the SVD therefore identifies the dominant directions of parameter change that occurred while optimizing that loss. Orthogonality is enforced only with respect to this data-driven subspace, not with respect to an analytically derived pixel-loss Hessian. The manuscript does not contain a formal derivation showing alignment between the right singular vectors and the pixel objective, nor does it report an explicit ablation measuring how much pixel-metric degradation is prevented by the projection. We will revise the manuscript to (i) add a short derivation clarifying that the subspace is empirical rather than theoretically guaranteed to be loss-aligned and (ii) include an ablation that quantifies the change in pixel metrics when semantic gradients are projected versus unprojected. These additions will be placed in the method section and supplementary material. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity; OGP is an explicit algorithmic construction

full rationale

The paper defines FreqMoE via a non-parametric FFT router and OGP via an explicit SVD-plus-nullspace projection on expert weight deltas; the orthogonality claim follows directly from linear algebra on the constructed subspace rather than from any fitted parameter renamed as a prediction or from a self-citation chain. No equation or result is shown to equal its own input by construction, and the central guarantee is presented as a property of the proposed procedure itself. The derivation is therefore self-contained.

Assumptions & free parameters 0 free parameters · 2 assumptions · 0 invented entities

Abstract-only review yields no explicit free parameters or invented entities; the approach relies on standard linear-algebra primitives (SVD, FFT) treated as domain assumptions and on the unstated premise that LoRA experts can be specialized by frequency signatures.

assumptions (2)

domain assumption SVD on combined expert weight deltas isolates a meaningful pixel-fidelity subspace
Invoked in the description of OGP to define the projection target.
domain assumption FFT-based features provide stable, interpretable signatures of real-world corruption types
Used to justify non-parametric routing in FreqMoE.

how reviews work

0 comments

Cite this review

Pith. "Pith review of FreqOrtho-SR: Frequency-Guided Orthogonal Expert Learning for Real-World Image Super-Resolution." pith.science (2026). https://pith.science/paper/QJVPDDPN

@misc{pith2026260628745,
  author       = {Pith},
  title        = {Pith review of: FreqOrtho-SR: Frequency-Guided Orthogonal Expert Learning for Real-World Image Super-Resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QJVPDDPN}},
  note         = {Machine review of arXiv:2606.28745}
}

abstract

Diffusion prior-based methods have shown impressive results in real-world image super-resolution (ISR), yet two key challenges persist: balancing pixel-level fidelity with semantic quality, and adapting to diverse degradations. Existing dual-branch approaches freeze the pixel module during semantic training, but the semantic branch can still expand capacity within the pixel subspace, precluding genuine perceptual improvement. Moreover, using a single static adapter cannot generalize across heterogeneous real-world corruptions. To address both issues, we propose FreqOrtho-SR, which comprises: $\textbf{Freq}$uency-guided Mixture of LoRA Experts (FreqMoE), it routes inputs to specialized experts via a non-parametric FFT-based degradation-feature extractor that encodes frequency-domain signatures, enabling stable and interpretable specialization across corruption types; and $\textbf{Ortho}$gonal Gradient Projection (OGP), which reframes the dual-objective optimization as a subspace-constrained problem: by extracting the pixel-fidelity subspace via SVD on combined expert weight deltas and projecting semantic gradients onto its null space, OGP guarantees orthogonality between the two objectives, enabling genuinely complementary learning without mutual interference. Experiments show that FreqOrtho-SR achieves competitive overall performance and a strong fidelity-perception trade-off across multiple benchmarks with efficient single-step inference. The source code of our method can be found at $\href{https://github.com/sonhm3029/FreqOrtho-SR}{\texttt{sonhm3029/FreqOrtho-SR}}$.

Figures

Figures reproduced from arXiv: 2606.28745 by the authors.

**Figure 1.** Visual and quantitative comparison on DRealSR. Values below each patch denote PSNR↑/LPIPS↓, with best in red. Our FreqOrtho-SR achieves the best fidelity and perceptual quality among one-step diffusion methods. inherent trade-off between perception and distortion [2]. While pixel-level regression methods using L1 or L2 losses achieve high fidelity metrics (e.g., PSNR, SSIM), their results are perceptually unsatisfy… view at source ↗

**Figure 2.** Overall architecture of our FreqOrtho-SR. The FreqMoE and Semantic LoRA modules are optimized for pixel-level and semantic-level enhancements, respectively. FreqMoE utilizes the power of Mixture of LoRA Experts to adaptively handle diverse degradation types via FFT-based gating, while Orthogonal Gradient Projection ensures the semantic LoRA learns in a subspace orthogonal to the pixel-level representations, reducing… view at source ↗

**Figure 3.** (a) FFT magnitude spectra (log scale) of LR images under different degradations. Blur suppresses high-frequency energy (dim outer ring), noise raises it uniformly (bright outer ring), and JPEG introduces periodic block artifacts (grid pattern). These signatures enable our non-parametric extractor to reliably identify degradation types. (b) Structure of the FreqMoE module which routes inputs to specialized LoRA expe… view at source ↗

Figures from the paper (8 more)

**Figure 4.** Figure 4: Visual comparisons of one-step DM-based SR methods. Zoom in for best view. across metrics. In contrast, FreqOrtho-SR achieves the best or second-best in several cases, indicating substantial improvements in the real-world ISR field. Qualitative results [PITH_FULL_IMAG…

**Figure 5.** Figure 5: Per-layer projection energy ratio onto the pixel-level subspace. Subspace orthogonality analysis. To directly verify that OGP enforces subspace independence and explain the perceptual gains when adding OGP to FreqMoE in [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]

**Figure 6.** Figure 6: Qualitative adjustable SR results with different pixel-semantic guidance scales. The vertical axis varies the pixel-level guidance scale λpix, while the horizontal axis varies the semantic-level guidance scale λsem. Qualitative analysis [PITH_FULL_IMAGE:figures/full_f…

**Figure 7.** Figure 7: Pearson correlation between degradation features and expert routing probabilities in FreqMoE, computed over 150 RealESRGAN degradation runs. Each heatmap corresponds to a different UNet component (encoder, decoder, others). Rows represent degradation features: radial …

**Figure 8.** Figure 8 [PITH_FULL_IMAGE:figures/full_fig_p030_8.png]

**Figure 9.** Figure 9: Additional visual comparisons between FreqOrtho-SR and competing methods from the main paper. Please zoom in for a better view [PITH_FULL_IMAGE:figures/full_fig_p033_9.png]

**Figure 10.** Figure 10: Visual comparisons between FreqOrtho-SR and GAN-based SR methods. Please zoom in for a better view [PITH_FULL_IMAGE:figures/full_fig_p034_10.png]

**Figure 11.** Figure 11: Visual comparisons between FreqOrtho-SR and multi-step SR methods. Please zoom in for a better view [PITH_FULL_IMAGE:figures/full_fig_p035_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 14 canonical work pages

[1]

In: International Conference on Artificial Intelligence and Statis- tics

Farajtabar, M., Azizan, N., Mott, A., Li, A.: Orthogonal gradient descent for con- tinual learning. In: International Conference on Artificial Intelligence and Statis- tics. pp. 3762–3773 (2020)

work page 2020
[2]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Liang, J., Zeng, H., Zhang, L.: Details or artifacts: A locally discriminative learning approach to realistic image super-resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5657–5666 (2022)

work page 2022
[3]

In: European conference on computer vision

Lin, X., He, J., Chen, Z., Lyu, Z., Dai, B., Yu, F., Qiao, Y., Ouyang, W., Dong, C.: Diffbir: Toward blind image restoration with generative diffusion prior. In: European conference on computer vision. pp. 430–448. Springer (2024)

work page 2024
[4]

Advances in neural information processing systems30(2017)

Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. Advances in neural information processing systems30(2017)

work page 2017
[5]

arXiv preprint arXiv:2103.09762 (2021)

Saha, G., Garg, I., Roy, K.: Gradient projection memory for continual learning. arXiv preprint arXiv:2103.09762 (2021)

work page arXiv 2021
[6]

Sun, L., Wu, R., Ma, Z., Liu, S., Yi, Q., Zhang, L.: Pixel-level and semantic-level adjustablesuper-resolution:Adual-loraapproach.In:ProceedingsoftheComputer Vision and Pattern Recognition Conference. pp. 2333–2343 (2025)

work page 2025
[7]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Wang, S., Li, X., Sun, J., Xu, Z.: Training networks in null space of feature co- variance for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 184–193 (2021)

work page 2021
[8]

In: Proceedings of the IEEE/CVF in- ternational conference on computer vision

Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In: Proceedings of the IEEE/CVF in- ternational conference on computer vision. pp. 1905–1914 (2021)

work page 1905

Show all 14 references

[9]

Advances in Neural Information Processing Systems 37, 92529–92553 (2024)

Wu, R., Sun, L., Ma, Z., Zhang, L.: One-step effective diffusion network for real- world image super-resolution. Advances in Neural Information Processing Systems 37, 92529–92553 (2024)

2024
[10]

In: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition

Wu, R., Yang, T., Sun, L., Zhang, Z., Li, S., Zhang, L.: Seesr: Towards semantics- aware real-world image super-resolution. In: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition. pp. 25456–25467 (2024)

2024
[11]

In: European conference on computer vision

Yang, T., Wu, R., Ren, P., Xie, X., Zhang, L.: Pixel-aware stable diffusion for real- istic image super-resolution and personalized stylization. In: European conference on computer vision. pp. 74–91. Springer (2024)

2024
[12]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Yu,F.,Gu,J.,Li,Z.,Hu,J.,Kong,X.,Wang,X.,He,J.,Qiao,Y.,Dong,C.:Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 25669–25680 (2024)

2024
[13]

Nature Machine Intelligence1(8), 364–372 (2019)

Zeng, G., Chen, Y., Cui, B., Yu, S.: Continual learning of context-dependent pro- cessing in neural networks. Nature Machine Intelligence1(8), 364–372 (2019)

2019
[14]

In: Proceedings of the IEEE/CVF international conference on computer vision

Zhang, K., Liang, J., Van Gool, L., Timofte, R.: Designing a practical degradation model for deep blind image super-resolution. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4791–4800 (2021)

2021

Pith tools

Reviewed June 30, 2026 · model on record in the stance chip above.

[1] [1]

In: International Conference on Artificial Intelligence and Statis- tics

Farajtabar, M., Azizan, N., Mott, A., Li, A.: Orthogonal gradient descent for con- tinual learning. In: International Conference on Artificial Intelligence and Statis- tics. pp. 3762–3773 (2020)

work page 2020

[2] [2]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Liang, J., Zeng, H., Zhang, L.: Details or artifacts: A locally discriminative learning approach to realistic image super-resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5657–5666 (2022)

work page 2022

[3] [3]

In: European conference on computer vision

Lin, X., He, J., Chen, Z., Lyu, Z., Dai, B., Yu, F., Qiao, Y., Ouyang, W., Dong, C.: Diffbir: Toward blind image restoration with generative diffusion prior. In: European conference on computer vision. pp. 430–448. Springer (2024)

work page 2024

[4] [4]

Advances in neural information processing systems30(2017)

Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. Advances in neural information processing systems30(2017)

work page 2017

[5] [5]

arXiv preprint arXiv:2103.09762 (2021)

Saha, G., Garg, I., Roy, K.: Gradient projection memory for continual learning. arXiv preprint arXiv:2103.09762 (2021)

work page arXiv 2021

[6] [6]

Sun, L., Wu, R., Ma, Z., Liu, S., Yi, Q., Zhang, L.: Pixel-level and semantic-level adjustablesuper-resolution:Adual-loraapproach.In:ProceedingsoftheComputer Vision and Pattern Recognition Conference. pp. 2333–2343 (2025)

work page 2025

[7] [7]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Wang, S., Li, X., Sun, J., Xu, Z.: Training networks in null space of feature co- variance for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 184–193 (2021)

work page 2021

[8] [8]

In: Proceedings of the IEEE/CVF in- ternational conference on computer vision

Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In: Proceedings of the IEEE/CVF in- ternational conference on computer vision. pp. 1905–1914 (2021)

work page 1905