acceptodds
Under review as a conference paper at ICLR 2027

PhaDiSR: Phase-Aware One-Step Diffusion Transformer for Real-World Image Super-Resolution

Abstract

Real-world image super-resolution increasingly benefits from the strong generative priors of diffusion transformers (DiTs), yet one-step DiT super-resolution remains prone to periodic, grid-like artifacts. Using a translation-equivariance diagnostic, we trace these artifacts to token-phase dependence: prediction discrepancies in packed DiTs vary periodically with the packing stride, whereas a convolutional U-Net counterpart shows no such periodicity. Motivated by this diagnosis, we propose PhaDiSR, a phase-aware one-step diffusion transformer for real-world image super-resolution that addresses token-phase dependence through velocity correction and wavelet discrimination. At the prediction level, Polyphase Token Reassembly exposes the implicit token phases in the packed velocity prediction, models their local spatial relations, and reassembles a phase-aware correction before the endpoint update. To penalize the resulting artifacts, Token-Phase Wavelet Discriminator uses multi-scale, multi-shift Haar supervision to effectively capture diverse local token-phase groupings. To facilitate training, Self-Prior Projection uses the frozen initialization prior to construct projected training references. PhaDiSR eliminates periodic grid artifacts and achieves leading perceptual fidelity across standard benchmarks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Diffuison model, Diffusion Transformer, Image super-resolution

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_GAmvOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the problem of periodic grid artifacts in one-step Diffusion Transformer (DiT) models for real-world image super-resolution. Through a translation equivariance analysis, the authors investigate the phenomenon of token-phase dependence. To mitigate this issue, they propose PhaDiSR, which consists of Polyphase Token Reassembly (PTR) for velocity prediction correction, a Token-Phase Wavelet Discriminator (TPWD) based on multi-scale and multi-shift Haar transforms, and Self-Prior Projection (SPP) for constructing training references. Experimental results demonstrate that the proposed method effectively alleviates periodic artifacts and achieves competitive perceptual quality across multiple real-world image super-resolution benchmarks.

Critical strengths

1.The paper addresses the practically relevant problem of periodic artifacts in one-step DiT-based super-resolution. The motivation is clear, and the problem has practical significance.

2.The proposed method is thoughtfully designed to address the identified problem. PTR performs phase-aware correction in the velocity prediction space, while TPWD provides supervision in the image space. These two components complement each other in mitigating periodic artifacts.

3.The experimental evaluation is reasonably comprehensive, including quantitative and qualitative comparisons on multiple datasets, as well as ablation studies of the main components. The proposed method achieves competitive results on several no-reference perceptual quality metrics.

Critical weaknesses

1.OP4KSR[1] has already provided a detailed analysis of the causes of periodic grid artifacts, showing that they are related not only to latent token packing but also to positional encoding, such as RoPE. This prior work somewhat weakens the novelty of the problem analysis presented in this paper. The authors should further clarify the differences between PhaDiSR and OP4KSR and highlight the specific advantages of the proposed approach.

2.Conventional GAN-based adversarial supervision can already suppress periodic grid artifacts to some extent. Therefore, it remains unclear to what extent TPWD alone, without PTR, can suppress these artifacts, and how much additional improvement TPWD provides over a conventional GAN discriminator. More direct quantitative and qualitative comparisons would help establish the effectiveness and necessity of the proposed TPWD design.

3.The paper lacks detailed inference efficiency comparisons. Since PTR introduces additional computational overhead, the authors should report inference latency, GPU memory usage, and FLOPs compared with the baseline and existing methods under identical settings.

[1]Chengyan Deng et al. OP4KSR: One-Step Patch-Free 4K Super-Resolution with Periodic Artifact Suppression

Questions for authors

Please see the three points raised in the Weaknesses section.

Reviewer_3obvOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies periodic grid artifacts in one-step DiT-based real-world image super-resolution. The authors observe that packed DiTs exhibit a translation-equivariance error whose spatial periodicity follows the latent packing stride, and refer to this behavior as token-phase dependence.

Based on this observation, the paper proposes PhaDiSR with three main components. Polyphase Token Reassembly (PTR) explicitly separates the phases within packed velocity predictions and models intra-cell and cross-boundary interactions before the one-step endpoint update. Token-Phase Wavelet Discriminator (TPWD) introduces multi-scale and multi-shift Haar-based adversarial supervision to penalize phase-dependent artifacts in image space. Self-Prior Projection (SPP) maps the original HQ training references through the frozen initialization prior before synthesizing the corresponding degraded inputs.

The paper evaluates the method on paired RealSR, DrealSR, and ScreenSR benchmarks and several unpaired real-world datasets, and reports additional component, mechanism, and cross-backbone experiments.

Critical strengths
  1. The diagnostic perspective on periodic artifacts is interesting. The translation-equivariance experiment provides a concrete way to characterize an artifact that is otherwise often treated only qualitatively. In particular, the observation that the mismatch follows the packing stride for packed DiTs, while a U-Net counterpart does not exhibit a comparable pattern, provides a plausible architectural explanation for the grid artifacts.

  2. The design of PTR and TPWD is reasonably connected to the diagnosed problem. PTR operates directly on the phase structure induced by packing and performs correction before the endpoint update, while TPWD attempts to supervise related discontinuities in image space using shifted Haar groupings. The two components therefore have a clearer connection to the proposed failure mode than a generic additional restoration loss would have.

  3. The experimental scope is relatively broad. The paper evaluates both paired and unpaired real-world benchmarks, compares against several recent U-Net and DiT restoration methods, provides component analyses, and also tests the framework on multiple DiT foundations. PhaDiSR achieves particularly strong results on several no-reference perceptual metrics and A-FINE, indicating that the complete system is competitive.

Critical weaknesses
  1. The current experiments do not sufficiently isolate the contribution of the proposed phase-aware design. The main SOTA table compares complete systems such as TSD-SR, ODTSR, StrSR, and PhaDiSR, which use different backbones, training data, losses, and adaptation strategies. Therefore, these results demonstrate that the complete PhaDiSR system is competitive, but do not establish that its proposed one-step phase-aware strategy is itself superior. A key missing experiment is a matched-backbone comparison under the same training data, Real-ESRGAN degradation, timestep, optimization budget, and trainable scope, e.g., a vanilla one-step FLUX.2 [klein] baseline versus the same backbone equipped with PhaDiSR. The cross-backbone experiments show that the framework can be applied to several DiT foundations, but they do not provide backbone-only versus backbone+PhaDiSR pairs, so the claim that the method universally improves different foundations is not directly supported.
    There is also an incomplete causal link between the proposed diagnostic and the solution. The paper introduces Ef(δ)E_f(\delta) as the central measure of token-phase dependence, but does not report how this quantity changes after applying PTR, TPWD, or the full model. Direct before/after measurements of Ef(δ)E_f(\delta), together with an image-domain periodicity measure, would be more convincing than relying mainly on general image-quality metrics. In addition, w/o TPWD appears to remove the adversarial supervision altogether, so it does not isolate the benefit of the proposed phase-aware wavelet discriminator from that of using adversarial supervision in general. The internal PTR studies are also inference-time pathway removals rather than clearly retrained architectural ablations.

  2. The novelty and evaluation of SPP, as well as the recent related-work coverage, need clearer justification. The broader idea behind SPP has close precedents that are not discussed. Relevant examples include Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon, “SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations,” ICLR 2022; Tianshu Kuai, Sina Honari, Igor Gilitschenski, and Alex Levinshtein, “Towards Unsupervised Blind Face Restoration using Diffusion Prior,” WACV 2025; and Yuyang Hu, Mojtaba Sahraee-Ardakan, Arpit Bansal, Kangfu Mei, Chenyang Qi, Peyman Milanfar, and Mauricio Delbracio, “Generative Manifold Distillation: Aligning Restoration Trajectories with Natural Image Prior,” ECCV 2026.

SPP is not identical to these methods—the specific use of a small-noise self-projection of an existing HQ reference followed by re-degradation may still be novel—but the paper should distinguish this specific contribution from the broader existing idea of generative-prior projection and pseudo-target construction. The phrase “Better than GT” in Fig. 3 is also stronger than the evidence supports. The experiment only shows that the projected references obtain higher scores on several no-reference IQA models; this does not necessarily imply that they are perceptually superior to the original GT according to human observers or that all fine structures are preserved. This concern is relevant because PhaDiSR improves many no-reference perceptual metrics while not consistently improving PSNR/LPIPS/DISTS, and the SPP ablation itself exhibits this fidelity-perception trade-off.
Some qualitative results raise a similar concern: in the electronics/breadboard example of Fig. 7, PhaDiSR introduces a specific local dark-dot structure that is not clearly supported by the LQ input, while no GT is provided to verify whether this is faithful recovery or hallucination.
Finally, the comparison does not cover several recent one-step SR methods available before submission, including Tianyi Zhang et al., “Time-Aware One Step Diffusion Network for Real-World Image Super-Resolution,” CVPR 2026; Jonghee Back et al., “IFCSR: Inference-Free Fidelity-Realism Control for One-Step Diffusion-based Real-World Image Super-Resolution,” CVPR 2026; Qiaosi Yi et al., “GDPO-SR: Group Direct Preference Optimization for One-Step Generative Image Super-Resolution,” CVPR 2026; and Hao Chen et al., “Bridging Fidelity-Reality with Controllable One-Step Diffusion for Image Super-Resolution,” CVPR 2026. FiDeSR and VOSR are cited by the paper, but FiDeSR is not included in the main comparison, and VOSR is not compared on ScreenSR despite its close connection to that benchmark. Several recent references also remain cited only through outdated arXiv versions.

Questions for authors
  1. Can the authors provide a controlled comparison under the same backbone and training protocol? At minimum, please compare a vanilla one-step FLUX.2 [klein] baseline with PhaDiSR under identical training data, Real-ESRGAN degradation, timestep, trainable scope, optimization budget, and comparable loss settings. If existing one-step strategies can reasonably be adapted to the same backbone, a matched comparison would further clarify whether the gain comes from the proposed strategy rather than differences between complete systems. For the cross-backbone experiments, please also provide backbone-only versus backbone+PhaDiSR results where possible.

  2. Does PhaDiSR directly reduce the token-phase dependence measured by the paper’s own diagnostic? Please report Ef(δ)E_f(\delta) for the original one-step backbone, +PTR, +TPWD, and the complete model. A direct image-domain periodicity measure would also help verify that the specific artifact identified in Section 3.1 is actually reduced. It would also be useful to clarify how token-phase dependence differs from other explanations of periodic artifacts discussed in recent work.

  3. Can the authors provide cleaner controls for PTR and TPWD? For TPWD, does w/o TPWD remove adversarial supervision entirely? If so, please compare against a conventional capacity-matched discriminator and separately ablate the Haar and multi-shift components. For PTR, the internal analyses appear to be inference-time pathway removals; can the authors provide retrained variants without the DC, contrast, or boundary components?

  4. Can the authors clarify the novelty of SPP relative to prior generative-prior-based projection and pseudo-target construction? In particular, please discuss Chenlin Meng et al., “SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations,” ICLR 2022; Tianshu Kuai et al., “Towards Unsupervised Blind Face Restoration using Diffusion Prior,” WACV 2025; and Yuyang Hu et al., “Generative Manifold Distillation: Aligning Restoration Trajectories with Natural Image Prior,” ECCV 2026. Is the main novelty of SPP specifically the projection of existing HQ references through the same initialization prior followed by re-degradation?

  5. Can the authors provide stronger evidence for the “Better than GT” interpretation of SPP? The current evidence shows higher scores on several no-reference IQA models, but these metrics are not equivalent to human preference or guaranteed content fidelity. A human preference study or analysis of semantic/fine-structure preservation would make this claim more convincing. Relatedly, in the electronics/breadboard example in Fig. 7, PhaDiSR appears to introduce a specific local dark-dot structure that is not clearly supported by the LQ input; can the authors provide the corresponding GT/reference or source dataset to verify whether this is faithful recovery rather than hallucination?

  6. Why are several recent one-step SR methods absent from the comparison? Please discuss Tianyi Zhang et al., “Time-Aware One Step Diffusion Network for Real-World Image Super-Resolution,” CVPR 2026; Jonghee Back et al., “IFCSR: Inference-Free Fidelity-Realism Control for One-Step Diffusion-based Real-World Image Super-Resolution,” CVPR 2026; Qiaosi Yi et al., “GDPO-SR: Group Direct Preference Optimization for One-Step Generative Image Super-Resolution,” CVPR 2026; and Hao Chen et al., “Bridging Fidelity-Reality with Controllable One-Step Diffusion for Image Super-Resolution,” CVPR 2026. FiDeSR and VOSR are already cited, but why are they not included in the corresponding main comparisons, particularly VOSR on ScreenSR?

  7. What exact ODTSR configuration was used for the reported comparison? Since ODTSR is controllable and different fidelity settings can lead to different fidelity/perception trade-offs, please specify the checkpoint, fidelity setting, prompt configuration, and preprocessing used to obtain the reported results.

3 comments
AuthorsOpenReview ↗

Thank you for your insightful review and kind follow up

Dear Reviewer 3obv,

Thank you again for your thorough and insightful review, which helped us substantially strengthen the paper. As the discussion period closes soon, we wanted to kindly check if our responses, revised manuscript, and additional experiments have addressed all of your concerns. Please let us know if any further clarification would be helpful. We truly appreciate your time and consideration!

Thank you for the detailed response. The new matched-backbone comparison, token-phase measurement, discriminator control, retrained PTR ablations, and additional baselines address most of my original concerns.

My main remaining concern is novelty. The response makes it clearer that SPP is an SDEdit-style projection adapted to paired SR training. I think this is useful, but still relatively incremental. The new results also suggest that the gains are more consistent on perceptual metrics than on fidelity metrics.

Given the additional evidence and revisions, I have increased my score.

Thank you very much for your thoughtful response, sharp perspective, and decision to raise your score! Your constructive guidance throughout the review process has helped us sharpen the positioning and rigor of our paper, and we will ensure these discussions are further refined in the final version.

Reviewer_Q86DOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PhaDiSR addresses periodic grid artifacts in super-resolution with a one-step diffusion Transformer. The authors first study how these artifacts relate to token phase. They then use PTR to coordinate velocity predictions within grid cells and across cell boundaries, and TPWD for multi-scale, multi-translation wavelet supervision. SPP generates the reference images used for training. The authors tested the method on multiple real-world datasets and several DiT backbones. The main gains were in perceptual quality.

Critical strengths
  1. The approach flows logically from problem analysis to method design. Translation experiments show that token phase affects predictions. PTR addresses this issue through velocity prediction, while TPWD addresses it through image supervision.
  2. The tests cover both paired and unpaired datasets. The method shows consistent gains on perceptual metrics and is also competitive on some full-reference metrics.
  3. The appendix provides quantitative results for response window localization, gradient intervention, and PTR intervention during inference. These results help clarify each module’s role beyond what the images alone show. Inference still takes one step and does not require a discriminator.
Critical weaknesses
  1. The paper uses translation inconsistency to diagnose the problem, but it does not report how much this inconsistency drops after correction on the same backbone. It also does not test fixed and multi-translation supervision separately.
  2. The cross-backbone experiments mainly report final performance. Without a comparison under the same training conditions, it is hard to tell whether every backbone benefits from PhaDiSR.
  3. The pre-trained DiT in Figure 2 looks like a U-Net, which does not match the Transformer backbone described in the text. We recommend showing the actual Transformer modules used.
Questions for authors
  1. Could you show the translation inconsistency curves for the base model, the full method, and key ablation variants side by side? How much do the results differ between fixed and multi-translation supervision?
  2. Could you compare several representative backbones with and without PhaDiSR under the same training conditions? How much does each component contribute?
  3. For text, regular geometric structures, and repetitive textures, how well are the original details preserved when perceptual quality improves? This might explain why perceptual quality and distortion metrics do not always improve together across datasets.
  4. The text says that validation fidelity is best at α=1\alpha=1, but Figure 11(a) seems to show a larger PSNR gain at α=1.25\alpha=1.25. Is there an error in the text or the figure, or does “best” refer to a different metric?
1 comment

Thank you for the reply. I have no further questions. I will maintain my original rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.