Pith. sign in

REVIEW 3 major objections 4 minor 62 references

One-Step Diffusion-based Real-World Image Super-Resolution with Visual Perception Distillation

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read VPD-SR claims that a one-step diffusion super-resolution model, trained to match a 15-step teacher while aligning CLIP image embeddings and preserving high-frequency detail, can beat both the teacher and prior state-of-the-art methods on…

desk verdict Solid incremental one-step SR paper with a real metric-leak caveat: the CLIP semantic loss trains toward the same embedding space CLIPIQA evaluates, and the DWT loss is applied to latents without proving the latent-to-pixel frequency link. read the letter →

arxiv 2506.02605 v1 pith:BC2T72UL submitted 2025-06-03 cs.CV

Xue Wu , Jingwei Xin , Zhijun Tu , Jie Hu , Jie Li , Nannan Wang , Xinbo Gao This is my paper · ORCID
classification cs.CV
keywords imagesuper-resolutiondiffusionmodelone-stepdistillationknowledgeCLIPsemanticsupervisionhigh-frequencyperceptionlossdiscretewavelettransformreal-world
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 proposes VPD-SR, a method for compressing a multi-step diffusion-based image super-resolution model into a single-step model without sacrificing perceptual quality. It argues that previous one-step distillation methods lose semantic fidelity because they lack explicit semantic supervision, so it adds a loss that aligns the CLIP image embeddings of the restored image and the ground truth, a loss that preserves high-frequency detail via a discrete wavelet transform in the latent space, and an adversarial loss. The central claim is that with these additions one sampling step beats prior state-of-the-art methods and the 15-step teacher model on no-reference perceptual metrics, with best CLIPIQA of 0.683 on ImageNet-Test, 0.7635 on RealSR, and 0.7817 on RealSet65. If this holds, diffusion-based super-resolution can be made fast enough for practical use while keeping high visual quality.

What carries the argument

The central mechanism is the combination of three supervision signals added to vanilla distillation: (1) Explicit Semantic-aware Supervision, which uses the frozen CLIP image encoder to compute $L_{SD}=1-\cos(\mathbf{v}_{SR},\mathbf{v}_{GT})$, pulling the student's output embedding toward the ground-truth embedding and thereby enforcing semantic consistency without needing semantic annotations; (2) a High-Frequency Perception loss that applies a discrete wavelet transform to the VQGAN latent codes of teacher and student outputs and penalises the difference in the three high-frequency sub-bands (horizontal, vertical, diagonal); and (3) an adversarial loss with a patch-based discriminator operating on latent codes. These losses are combined as $L_{f_\theta}=L_{\text{distill}}+\lambda_1 L_{HFP}+\lambda_2 L_{SD}+\lambda_3 L_{adv}$, and the whole objective is computed in the VQGAN latent space defined by the frozen encoder and decoder.

What would settle it

Retrain VPD-SR with the high-frequency perception loss computed on the decoded pixel-space images instead of on the latent codes, keeping all other losses and hyperparameters fixed; if the CLIPIQA and MUSIQ gains over the vanilla-distillation baseline persist, the latent-space version of the loss is not the active ingredient. Alternatively, compare the Fourier spectra of the student outputs with and without the high-frequency loss to check whether the loss actually increases the high-frequency content of the final image.

Watch

Extended reading notes

Core claim

VPD-SR distills the multi-step ResShift teacher into a one-step student by initializing the student from the teacher and using the teacher's final-step output as the distillation target. On top of the vanilla distillation loss, the method adds explicit semantic-aware supervision: the student's decoded output and the ground-truth image are passed through a frozen CLIP image encoder, and the loss $L_{SD}=1-\cos(\mathbf{v}_{SR},\mathbf{v}_{GT})$ between the two embeddings is minimized. It also adds a high-frequency perception loss $L_{HFP}$ that applies a discrete wavelet transform to the teacher and student latent codes and penalizes differences in the horizontal, vertical, and diagonal high-frequency sub-bands, and an adversarial loss from a patch-based discriminator on the VQGAN latent codes. The paper claims that this combination produces a single-step model that outperforms previous state-of-the-art methods and the teacher model itself on CLIPIQA and MUSIQ across synthetic and real-world datasets, while running about ten times faster than the teacher; it also acknowledges that the advantage is concentrated in no-reference metrics and that full-reference fidelity is a minor limitation.

Load-bearing premise

The paper never shows that the high-frequency detail it preserves inside the compressed representation is the same high-frequency detail that its pixel-space analysis shows drives perceptual quality; if those differ, the high-frequency loss may not be the active ingredient.

Editorial extensions

If this is right

  • A single sampling step can match or exceed the perceptual quality of a 15-step diffusion teacher, cutting inference time from about 0.68 s to about 0.058 s on an A100 for 64-to-256 super-resolution.
  • CLIP-based semantic supervision can replace explicit semantic cues extracted from degraded low-resolution inputs, which the paper argues are unreliable for real-world images.
  • High-frequency detail is identified as the main driver of perceptual-quality improvement during diffusion SR inference, so explicitly preserving it during distillation transfers that gain to the one-step model.
  • On CLIPIQA the one-step model also surpasses multi-step Stable Diffusion-based SR methods on RealSR and DRealSR, although some multi-step methods retain higher MUSIQ and MANIQA scores.

Reading between the lines

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

  • Because CLIPIQA itself is computed using a CLIP image encoder, the CLIP-embedding training loss may inflate the reported CLIPIQA numbers specifically; the smaller gains on MUSIQ are consistent with this possibility, and a fair comparison would include a version trained without the CLIP loss using a non-CLIP perceptual metric.
  • The high-frequency perception loss is applied to VQGAN latent codes, while the motivating Fourier analysis is done on decoded pixel images; testing the DWT loss on decoded images instead would reveal whether the latent-space high-frequency bands actually track perceptually relevant image frequencies.
  • The ablation table shows the adversarial loss alone raises CLIPIQA from 0.6548 to 0.6830 on ImageNet-Test, so part of the one-step gain may come primarily from adversarial training rather than from the new semantic and high-frequency losses.
  • The same recipe of semantic embedding alignment, high-frequency preservation, and adversarial training could in principle transfer to other one-step distillation problems such as text-to-image or video super-resolution, where semantic fidelity and high-frequency detail are also bottlenecks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper introduces VPD-SR, a one-step diffusion-based image super-resolution method obtained by distilling a pre-trained 15-step ResShift teacher into a single-step student. The student is trained with a vanilla distillation loss, a DWT-based high-frequency perception loss applied to VQGAN latent codes, a CLIP embedding cosine-similarity semantic loss, and an adversarial loss. Experiments on ImageNet-Test, RealSR*, RealSet65, RealSR†, and DRealSR report state-of-the-art CLIPIQA scores, competitive MUSIQ and NIQE results, and a large inference speed-up over multi-step diffusion baselines. The ablation study in Table V shows that each loss component contributes to the reported metrics.

Significance. If the results hold, the paper is a useful contribution to efficient diffusion-based super-resolution: it shows a simple combination of semantic, high-frequency, and adversarial losses can condense a multi-step diffusion SR model into one step while retaining strong no-reference perceptual scores. The paper deserves explicit credit for a clearly organized ablation study (Table V) that demonstrates the incremental benefit of each loss, and for comparisons against a broad set of GAN, Transformer, and diffusion baselines. However, the significance is bounded by two issues. First, the headline quantitative claim rests heavily on CLIPIQA, a metric built on CLIP image features, while the semantic loss in Eq. (8) explicitly optimizes cosine similarity in that same embedding space, creating a partial train/eval alignment. Second, on the RealSR† and DRealSR datasets in Table III, VPD-SR is not the best method on MUSIQ or MANIQA, so the abstract's claim of "superior performance compared to both previous state-of-the-art methods" is overstated.

major comments (3)
  1. [Section III-C, Eq. (8), and Section IV-A] The semantic distillation loss L_SD minimizes 1 minus the cosine similarity between CLIP image embeddings of the student output and the ground-truth image, while CLIPIQA, the paper's primary no-reference metric, is a regressor trained on CLIP image features. The large CLIPIQA gains, particularly the jump from 0.6311 to 0.6548 on ImageNet-Test when L_SD is added in Table V, therefore reflect in part direct optimization of the evaluation metric's feature space rather than an independent measure of perceptual quality. The authors should either add a human perceptual study, use an IQA metric whose features do not come from CLIP, or clearly narrow the claim to CLIP-feature-aligned perceptual quality.
  2. [Section III-D, Eq. (9)] The high-frequency perception loss L_HFP is computed on VQGAN latent codes z_Tch and z_Stu, but the motivating evidence in Figures 2 and 3 is built from Fourier analysis of the teacher model's pixel-space predictions and their CLIPIQA trajectories. The manuscript never establishes that the DWT high-frequency sub-bands of latent codes correspond to perceptually relevant high-frequency image details. Without such validation, the CLIPIQA/MUSIQ improvements from L_HFP in Table V could be attributable to a generic regularization effect rather than to the intended high-frequency perception mechanism. Please provide a latent-space spectral analysis or an ablation with a pixel-space variant of the loss.
  3. [Abstract, Section IV-B, and Tables II/III] The abstract and Section IV-B claim that VPD-SR "achieves superior performance compared to both previous state-of-the-art methods and the teacher model," but Table III shows that on RealSR† and DRealSR, VPD-SR trails SeeSR and PASD on MUSIQ and MANIQA by substantial margins (for example, MUSIQ 66.38 vs. 69.77 on RealSR†, and MANIQA 0.5817 vs. 0.6487), while only CLIPIQA is consistently best. The claim should be restricted to the CLIPIQA metric and to the one-step setting, with a discussion of the mixed multi-metric results. In addition, all quantitative tables report single-run numbers without error bars; given that some CLIPIQA differences are small, the absence of variance information is a limitation.
minor comments (4)
  1. [Section I and Section III-B] There are multiple typos and grammatical errors, including "polt" instead of "plotted", "student moldel" instead of "student model", "the the predictive outputs", and "performance pf our proposed method". A careful proofread is needed.
  2. [Section III-D and Algorithm 1] The notation in Eq. (9) and Algorithm 1 is imprecise: the DWT sub-bands H, V, D are not explicitly defined as functions of the input latent code, and Algorithm 1 line 7 uses an expectation symbol that is unnecessary and potentially confusing. Please clarify the wavelet basis and decomposition level used in the DWT.
  3. [Section IV-A and Tables I/IV] The implementation details state that training uses an NVIDIA Tesla V100 GPU, while Table I reports runtime on an NVIDIA Tesla A100 GPU and Table IV reports runtime on a V100 GPU. The runtime comparison would be clearer if all measurements used the same GPU platform.
  4. [Section IV-C and Table V] The ablation text states that adding the high-frequency perception loss "significantly improves" CLIPIQA and MUSIQ, but the differences in some rows are small and no statistical significance test or repeated-run variance is reported. I recommend softening the language or adding error bars.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's results are empirical measurements driven by explicit training losses, not derivations that reduce to the evaluation metric by construction.

full rationale

VPD-SR makes no first-principles derivation whose conclusion is equivalent to its inputs. The training objective (Eq. 13) combines a distillation loss, a DWT high-frequency perception loss, a CLIP semantic loss (Eq. 8), and an adversarial loss; all are standard empirical loss functions. The CLIP semantic loss aligns student and ground-truth CLIP embeddings, and CLIPIQA is a learned regressor over CLIP features, so there is a real overlap between the optimized feature space and the primary no-reference metric. However, this is not an identity: Eq. (8) minimizes a cosine distance to the ground-truth embedding, while CLIPIQA is a separately trained quality regressor, so CLIPIQA gains are not forced by the loss by construction. The paper also reports MUSIQ, MANIQA, NIQE, and LPIPS, and its own Table III shows that VPD-SR does not win on MUSIQ and MANIQA on RealSR and DRealSR, indicating that the headline CLIPIQA advantage is not the only evidence and is not presented as a derived theorem. The Fourier and CLIPIQA trajectory analysis motivating the HFP loss is an empirical observation, not a circular premise. The only author-overlapping prior work cited, TAD-SR [33], is used solely as a comparison baseline and is not load-bearing for any claimed derivation. No self-citation chain, imported uniqueness theorem, or ansatz-smuggling is present. The paper is therefore self-contained as an empirical method paper, and any concern about metric alignment is a correctness or evaluation-design issue, not circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method rests on standard pretrained models (ResShift teacher, CLIP, VQGAN) and three hand-set loss weights. The main unstated premise is the pixel-space to latent-space transfer of high-frequency semantics. No new entities are invented.

free parameters (4)
  • lambda1 (HFP loss weight) = 0.1
    Set manually in Section IV-A; no sensitivity study reported.
  • lambda2 (semantic distillation loss weight) = 1.0
    Set manually in Section IV-A; no sensitivity study reported.
  • lambda3 (adversarial loss weight) = 0.1
    Set manually in Section IV-A; no sensitivity study reported.
  • DWT level and wavelet basis = not specified
    The HFP loss in Eq. (9) applies DWT to latent codes but the wavelet type and decomposition depth are not stated, leaving the loss dependent on an unspecified choice.
assumptions (4)
  • standard math The discrete wavelet transform is an invertible decomposition that separates high-frequency content from low-frequency content (used in Eq. (9)).
    Section III-D assumes DWT sub-bands H, V, D capture horizontal, vertical, and diagonal high-frequency information.
  • domain assumption Cosine similarity between CLIP image embeddings of the SR output and the GT image is a meaningful semantic/perceptual supervision signal.
    Section III-C defines LSD in Eq. (8) from cos in Eq. (7); if this similarity is a poor proxy, the semantic loss could distort outputs without improving true perception.
  • ad hoc to paper High-frequency components in the VQGAN latent space correspond to the high-frequency image details that improve perceptual quality.
    The motivation in Figures 2-3 is based on pixel-space prediction Fourier spectra, but Eq. (9) applies DWT to latent codes z; this transfer is asserted, not demonstrated.
  • domain assumption The pre-trained ResShift teacher with deterministic sampling provides a reliable distillation target and a good initialization.
    Section III-B and IV-A initialize the student from the teacher and distill the teacher's one-step output; errors in the teacher directly bound the student's ceiling.

how reviews work

0 comments
Cite this review

Pith. "Pith review of One-Step Diffusion-based Real-World Image Super-Resolution with Visual Perception Distillation." pith.science (2026). https://pith.science/paper/BC2T72UL

@misc{pith2026250602605,
  author       = {Pith},
  title        = {Pith review of: One-Step Diffusion-based Real-World Image Super-Resolution with Visual Perception Distillation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BC2T72UL}},
  note         = {Machine review of arXiv:2506.02605}
}
read the original abstract

Diffusion-based models have been widely used in various visual generation tasks, showing promising results in image super-resolution (SR), while typically being limited by dozens or even hundreds of sampling steps. Although existing methods aim to accelerate the inference speed of multi-step diffusion-based SR methods through knowledge distillation, their generated images exhibit insufficient semantic alignment with real images, resulting in suboptimal perceptual quality reconstruction, specifically reflected in the CLIPIQA score. These methods still have many challenges in perceptual quality and semantic fidelity. Based on the challenges, we propose VPD-SR, a novel visual perception diffusion distillation framework specifically designed for SR, aiming to construct an effective and efficient one-step SR model. Specifically, VPD-SR consists of two components: Explicit Semantic-aware Supervision (ESS) and High-Frequency Perception (HFP) loss. Firstly, the ESS leverages the powerful visual perceptual understanding capabilities of the CLIP model to extract explicit semantic supervision, thereby enhancing semantic consistency. Then, Considering that high-frequency information contributes to the visual perception quality of images, in addition to the vanilla distillation loss, the HFP loss guides the student model to restore the missing high-frequency details in degraded images that are critical for enhancing perceptual quality. Lastly, we expand VPD-SR in adversarial training manner to further enhance the authenticity of the generated content. Extensive experiments conducted on synthetic and real-world datasets demonstrate that the proposed VPD-SR achieves superior performance compared to both previous state-of-the-art methods and the teacher model with just one-step sampling.

Figures

Figures reproduced from arXiv: 2506.02605 by the authors.

Figure 1
Figure 1. Performance comparison on CLIPIQA (top) and MUSIQ (bottom) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. The CLIPIQA score trajectories for both the teacher model’s [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Overall framework of our proposed VPD-SR. We distill the given pre-trained teacher model into a fast one-step student model. To match the single-step [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Qualitative comparisons of different methods on two synthetic examples of the oed (b) SG(c) eSJG (d) SG(e) Sw() eSG(g) S [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Qualitative comparisons of different methods on two real examples of the [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Qualitative comparisons of different methods on four real examples of the [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

62 extracted references · 46 canonical work pages

  1. [1]

    Deep learning for image super- resolution: A survey,

    Z. Wang, J. Chen, and S. C. Hoi, “Deep learning for image super- resolution: A survey,” IEEE TPAMI, vol. 43, no. 10, pp. 3365–3387, 2020

    work page 2020

  2. [2]

    Multi-contrast mri arbitrary-scale super-resolution via dynamic implicit network,

    J. Wei, G. Yang, W. Wei, A. Liu, and X. Chen, “Multi-contrast mri arbitrary-scale super-resolution via dynamic implicit network,” IEEE Transactions on Circuits and Systems for Video Technology , pp. 1–1, 2025

    work page 2025

  3. [3]

    Wide weighted attention multi-scale network for accurate mr image super-resolution,

    H. Wang, X. Hu, X. Zhao, and Y . Zhang, “Wide weighted attention multi-scale network for accurate mr image super-resolution,” IEEE Transactions on Circuits and Systems for Video Technology , vol. 32, no. 3, pp. 962–975, 2022

    work page 2022

  4. [4]

    Frequency-assisted mamba for remote sensing image super-resolution,

    Y . Xiao, Q. Yuan, K. Jiang, Y . Chen, Q. Zhang, and C.-W. Lin, “Frequency-assisted mamba for remote sensing image super-resolution,” IEEE Transactions on Multimedia , vol. 27, pp. 1783–1796, 2025

    work page 2025

  5. [5]

    Local-global temporal difference learning for satellite video super- resolution,

    Y . Xiao, Q. Yuan, K. Jiang, X. Jin, J. He, L. Zhang, and C.-w. Lin, “Local-global temporal difference learning for satellite video super- resolution,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 4, pp. 2789–2802, 2023

    work page 2023

  6. [6]

    Video rescaling with recurrent diffusion,

    D. Li, Y . Liu, Z. Wang, and J. Yang, “Video rescaling with recurrent diffusion,” IEEE Transactions on Circuits and Systems for Video Tech- nology, vol. 34, no. 10, pp. 9386–9399, 2024

    work page 2024

  7. [7]

    Sequence unlearning for sequential recommender systems,

    S. Ye and J. Lu, “Sequence unlearning for sequential recommender systems,” in AI 2023: Advances in Artificial Intelligence, 2024, pp. 403– 415

    work page 2023

  8. [8]

    Cbash: Combined backbone and advanced selection heads with object semantic proposals for weakly supervised object detection,

    R. Xia, G. Li, Z. Huang, H. Meng, and Y . Pang, “Cbash: Combined backbone and advanced selection heads with object semantic proposals for weakly supervised object detection,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 10, pp. 6502–6514, 2022

    work page 2022

Show all 62 references
  1. [9]

    Robust recommender systems with rating flip noise,

    S. Ye and J. Lu, “Robust recommender systems with rating flip noise,” ACM Trans. Intell. Syst. Technol. , vol. 16, no. 1, pp. 1–19, 2024

    2024

  2. [10]

    Generative adversarial nets,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial nets,” in NeuralIPS, 2014, pp. 2672–2680

    2014

  3. [11]

    Details or artifacts: A locally discriminative learning approach to realistic image super-resolution,

    J. Liang, H. Zeng, and L. Zhang, “Details or artifacts: A locally discriminative learning approach to realistic image super-resolution,” in CVPR, 2022, pp. 5657–5666

    2022

  4. [12]

    Esrgan: Enhanced super-resolution generative adver- sarial networks,

    X. Wang, K. Yu, S. Wu, J. Gu, Y . Liu, C. Dong, Y . Qiao, and C. Change Loy, “Esrgan: Enhanced super-resolution generative adver- sarial networks,” in ECCVW, 2018, pp. 63–79

    2018

  5. [13]

    Real-esrgan: Training real- world blind super-resolution with pure synthetic data,

    X. Wang, L. Xie, C. Dong, and Y . Shan, “Real-esrgan: Training real- world blind super-resolution with pure synthetic data,” in ICCV, 2021, pp. 1905–1914

    2021

  6. [14]

    Transformer for single image super-resolution,

    Z. Lu, J. Li, H. Liu, C. Huang, L. Zhang, and T. Zeng, “Transformer for single image super-resolution,” in CVPR, 2022, pp. 457–466

    2022

  7. [15]

    Pre-trained image processing transformer,

    H. Chen, Y . Wang, T. Guo, C. Xu, Y . Deng, Z. Liu, S. Ma, C. Xu, C. Xu, and W. Gao, “Pre-trained image processing transformer,” inCVPR, 2021, pp. 12 299–12 310

    2021

  8. [16]

    Learning attention from attention: Efficient self-refinement transformer for face super-resolution

    G. Li, J. Shi, Y . Zong, F. Wang, T. Wang, and Y . Gong, “Learning attention from attention: Efficient self-refinement transformer for face super-resolution.” in IJCAI, 2023, pp. 1035–1043

    2023

  9. [17]

    Freqformer: frequency-aware transformer for lightweight image super-resolution,

    T. Dai, J. Wang, H. Guo, J. Li, J. Wang, and Z. Zhu, “Freqformer: frequency-aware transformer for lightweight image super-resolution,” in IJCAI, 2024, pp. 731–739

    2024

  10. [18]

    Image super-resolution via iterative refinement,

    C. Saharia, J. Ho, W. Chan, T. Salimans, D. J. Fleet, and M. Norouzi, “Image super-resolution via iterative refinement,” IEEE TPAMI, vol. 45, no. 4, pp. 4713–4726, 2022

    2022

  11. [19]

    Srdiff: Single image super-resolution with diffusion probabilistic mod- els,

    H. Li, Y . Yang, M. Chang, S. Chen, H. Feng, Z. Xu, Q. Li, and Y . Chen, “Srdiff: Single image super-resolution with diffusion probabilistic mod- els,” Neurocomputing, vol. 479, pp. 47–59, 2022

    2022

  12. [20]

    Resdiff: Combining cnn and diffusion model for image super- resolution,

    S. Shang, Z. Shan, G. Liu, L. Wang, X. Wang, Z. Zhang, and J. Zhang, “Resdiff: Combining cnn and diffusion model for image super- resolution,” in AAAI, 2024, pp. 8975–8983

    2024

  13. [21]

    Come-closer-diffuse-faster: Accelerat- ing conditional diffusion models for inverse problems through stochastic contraction,

    H. Chung, B. Sim, and J. C. Ye, “Come-closer-diffuse-faster: Accelerat- ing conditional diffusion models for inverse problems through stochastic contraction,” in CVPR, 2022, pp. 12 413–12 422

    2022

  14. [22]

    Sdedit: Guided image synthesis and editing with stochastic differential equations,

    C. Meng, Y . He, Y . Song, J. Song, J. Wu, J.-Y . Zhu, and S. Ermon, “Sdedit: Guided image synthesis and editing with stochastic differential equations,” in ICLR, 2022

    2022

  15. [23]

    Resshift: Efficient diffusion model for image super-resolution by residual shifting,

    Z. Yue, J. Wang, and C. C. Loy, “Resshift: Efficient diffusion model for image super-resolution by residual shifting,” in NeurIPS, 2023

    2023

  16. [24]

    Exploiting diffusion prior for real-world image super-resolution,

    J. Wang, Z. Yue, S. Zhou, K. C. Chan, and C. C. Loy, “Exploiting diffusion prior for real-world image super-resolution,” IJCV, vol. 132, no. 12, pp. 5929–5949, 2024

    2024

  17. [25]

    Diffbir: Toward blind image restoration with generative diffusion prior,

    X. Lin, J. He, Z. Chen, Z. Lyu, B. Dai, F. Yu, Y . Qiao, W. Ouyang, and C. Dong, “Diffbir: Toward blind image restoration with generative diffusion prior,” in ECCV, 2024, pp. 430–448

    2024

  18. [26]

    Seesr: Towards semantics-aware real-world image super-resolution,

    R. Wu, T. Yang, L. Sun, Z. Zhang, S. Li, and L. Zhang, “Seesr: Towards semantics-aware real-world image super-resolution,” in CVPR, 2024, pp. 25 456–25 467

    2024

  19. [27]

    Pixel-aware stable dif- fusion for realistic image super-resolution and personalized stylization,

    T. Yang, R. Wu, P. Ren, X. Xie, and L. Zhang, “Pixel-aware stable dif- fusion for realistic image super-resolution and personalized stylization,” in ECCV, 2024, pp. 74–91

    2024

  20. [28]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” arXiv preprint arXiv:2010.02502 , 2020

    2010 arXiv

  21. [29]

    Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models,

    C. Lu, Y . Zhou, F. Bao, J. Chen, C. Li, and J. Zhu, “Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models,” arXiv preprint arXiv:2211.01095, 2022

    2022 arXiv

  22. [30]

    Progressive distillation for fast sampling of diffusion models,

    T. Salimans and J. Ho, “Progressive distillation for fast sampling of diffusion models,” arXiv preprint arXiv:2202.00512 , 2022

    2022 arXiv

  23. [31]

    Adversarial diffusion distillation,

    A. Sauer, D. Lorenz, A. Blattmann, and R. Rombach, “Adversarial diffusion distillation,” in ECCV, 2024, pp. 87–103

    2024

  24. [32]

    Sinsr: diffusion-based image super- resolution in a single step,

    Y . Wang, W. Yang, X. Chen, Y . Wang, L. Guo, L.-P. Chau, Z. Liu, Y . Qiao, A. C. Kot, and B. Wen, “Sinsr: diffusion-based image super- resolution in a single step,” in CVPR, 2024, pp. 25 796–25 805

    2024

  25. [33]

    One step diffusion-based super-resolution with time-aware distillation,

    X. He, H. Tang, Z. Tu, J. Zhang, K. Cheng, H. Chen, Y . Guo, M. Zhu, N. Wang, X. Gao et al., “One step diffusion-based super-resolution with time-aware distillation,” arXiv preprint arXiv:2408.07476 , 2024

    2024 arXiv

  26. [34]

    Addsr: Accelerating diffusion-based blind super-resolution with adversarial diffusion distillation,

    R. Xie, Y . Tai, C. Zhao, K. Zhang, Z. Zhang, J. Zhou, X. Ye, Q. Wang, and J. Yang, “Addsr: Accelerating diffusion-based blind super-resolution with adversarial diffusion distillation,” arXiv preprint arXiv:2404.01717, 2024

    2024 arXiv

  27. [35]

    One-step effective diffu- sion network for real-world image super-resolution,

    R. Wu, L. Sun, Z. Ma, and L. Zhang, “One-step effective diffu- sion network for real-world image super-resolution,” arXiv preprint arXiv:2406.08177, 2024

    2024 arXiv

  28. [36]

    Degradation-guided one-step image super-resolution with diffusion priors,

    A. Zhang, Z. Yue, R. Pei, W. Ren, and X. Cao, “Degradation-guided one-step image super-resolution with diffusion priors,” arXiv preprint arXiv:2409.17058, 2024

    2024 arXiv

  29. [37]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al., “Learning transferable visual models from natural language supervision,” in ICML. PMLR, 2021, pp. 8748–8763

    2021

  30. [38]

    Ssp- ir: Semantic and structure priors for diffusion-based realistic image restoration,

    Y . Zhang, H. Zhang, Z. Cheng, R. Xie, L. Song, and W. Zhang, “Ssp- ir: Semantic and structure priors for diffusion-based realistic image restoration,” IEEE Transactions on Circuits and Systems for Video Technology, pp. 1–1, 2025

    2025

  31. [39]

    Learning a deep convolutional network for image super-resolution,

    C. Dong, C. C. Loy, K. He, and X. Tang, “Learning a deep convolutional network for image super-resolution,” in ECCV, 2014, pp. 184–199

    2014

  32. [40]

    Multi-scale non-local attention network for image super-resolution,

    X. Wu, K. Zhang, Y . Hu, X. He, and X. Gao, “Multi-scale non-local attention network for image super-resolution,” Signal Processing , vol. 218, p. 109362, 2024

    2024

  33. [41]

    Swinir: Image restoration using swin transformer,

    J. Liang, J. Cao, G. Sun, K. Zhang, L. Van Gool, and R. Timofte, “Swinir: Image restoration using swin transformer,” in ICCV, 2021, pp. 1833–1844

    2021

  34. [42]

    Real-world super- resolution via kernel estimation and noise injection,

    X. Ji, Y . Cao, Y . Tai, C. Wang, J. Li, and F. Huang, “Real-world super- resolution via kernel estimation and noise injection,” in CVPRW, 2020, pp. 466–467

    2020

  35. [43]

    Designing a practical degradation model for deep blind image super-resolution,

    K. Zhang, J. Liang, L. Van Gool, and R. Timofte, “Designing a practical degradation model for deep blind image super-resolution,” in ICCV, 2021, pp. 4791–4800

    2021

  36. [44]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” in CVPR, 2022, pp. 10 684–10 695

    2022

  37. [45]

    Improved denoising diffusion proba- bilistic models,

    A. Q. Nichol and P. Dhariwal, “Improved denoising diffusion proba- bilistic models,” in ICML, 2021, pp. 8162–8171

    2021

  38. [46]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” in NeurIPS, 2020

    2020

  39. [47]

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps,

    C. Lu, Y . Zhou, F. Bao, J. Chen, C. Li, and J. Zhu, “Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps,” in NeurIPS, 2022

    2022

  40. [48]

    On distillation of guided diffusion models,

    C. Meng, R. Rombach, R. Gao, D. Kingma, S. Ermon, J. Ho, and T. Salimans, “On distillation of guided diffusion models,” in CVPR, 2023, pp. 14 297–14 306

    2023

  41. [49]

    You only need one step: Fast super-resolution with stable diffusion via scale distillation,

    M. Noroozi, I. Hadji, B. Martinez, A. Bulat, and G. Tzimiropoulos, “You only need one step: Fast super-resolution with stable diffusion via scale distillation,” in ECCV, 2024, pp. 145–161

    2024

  42. [50]

    Prolific- dreamer: High-fidelity and diverse text-to-3d generation with variational score distillation,

    Z. Wang, C. Lu, Y . Wang, F. Bao, C. Li, H. Su, and J. Zhu, “Prolific- dreamer: High-fidelity and diverse text-to-3d generation with variational score distillation,” in NeurIPS, 2023

    2023

  43. [51]

    Taming transformers for high- resolution image synthesis,

    P. Esser, R. Rombach, and B. Ommer, “Taming transformers for high- resolution image synthesis,” in CVPR, 2021, pp. 12 873–12 883. 11

    2021

  44. [52]

    Image-to-image translation with conditional adversarial networks,

    I. Phillip, Z. Jun-Yan, Z. Tinghui, A. Alexei et al. , “Image-to-image translation with conditional adversarial networks,” in CVPR, 2017, pp. 5967–5976

    2017

  45. [53]

    Efficient and degradation-adaptive network for real-world image super-resolution,

    J. Liang, H. Zeng, and L. Zhang, “Efficient and degradation-adaptive network for real-world image super-resolution,” in ECCV, 2022, pp. 574–591

    2022

  46. [54]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in CVPR. Ieee, 2009, pp. 248–255

    2009

  47. [55]

    Toward real-world single image super-resolution: A new benchmark and a new model,

    J. Cai, H. Zeng, H. Yong, Z. Cao, and L. Zhang, “Toward real-world single image super-resolution: A new benchmark and a new model,” in ICCV, 2019, pp. 3086–3095

    2019

  48. [56]

    Component divide-and-conquer for real-world image super-resolution,

    P. Wei, Z. Xie, H. Lu, Z. Zhan, Q. Ye, W. Zuo, and L. Lin, “Component divide-and-conquer for real-world image super-resolution,” in ECCV, 2020, pp. 101–117

    2020

  49. [57]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in CVPR, 2018, pp. 586–595

    2018

  50. [58]

    Exploring clip for assessing the look and feel of images,

    J. Wang, K. C. Chan, and C. C. Loy, “Exploring clip for assessing the look and feel of images,” in AAAI, 2023, pp. 2555–2563

    2023

  51. [59]

    Musiq: Multi-scale image quality transformer,

    J. Ke, Q. Wang, Y . Wang, P. Milanfar, and F. Yang, “Musiq: Multi-scale image quality transformer,” in ICCV, 2021, pp. 5148–5157

    2021

  52. [60]

    Maniqa: Multi-dimension attention network for no-reference image quality assessment,

    S. Yang, T. Wu, S. Shi, S. Lao, Y . Gong, M. Cao, J. Wang, and Y . Yang, “Maniqa: Multi-dimension attention network for no-reference image quality assessment,” in CVPR, 2022, pp. 1191–1200

    2022

  53. [61]

    A feature-enriched completely blind image quality evaluator,

    L. Zhang, L. Zhang, and A. C. Bovik, “A feature-enriched completely blind image quality evaluator,” IEEE TIP, vol. 24, no. 8, pp. 2579–2591, 2015. Xue Wu received the B.Sc. and M.Sc. degrees from Xi’an Polytechnic University, Xi’an, China, in 2021 and 2024, respectively. He is...

    2015

  54. [2021]

    He is currently a Lecturer with the State Key Laboratory of Integrated Services Networks, Xidian University. His research has been published in prestigious journals and academic venues, such as International Journal of Computer Vision (IJCV) and European Conference on Computer...

    2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.