pith. sign in

arxiv: 2604.19680 · v1 · submitted 2026-04-21 · 💻 cs.CV

IR-Flow: Bridging Discriminative and Generative Image Restoration via Rectified Flow

Zihao Fan , Xin Lu , Jie Xiao , Dong Li , Jie Huang , Xueyang Fu This is my paper

Pith reviewed 2026-05-10 02:42 UTC · model grok-4.3

classification 💻 cs.CV
keywords image restorationrectified flowgenerative modelsdiscriminative modelslinear transportderainingdenoisingfew-step sampling
0
0 comments X

The pith

Rectified Flow creates a direct linear path from degraded to clean images for fast and adaptable restoration.

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

The paper aims to establish that rectified flow models can bridge single-step discriminative restoration, which tends to average out fine details, and multi-step generative approaches, which are slow due to iterative sampling. It does this by building multilevel distribution flows across degradation levels and learning cumulative velocity fields that define straight transport trajectories to clean images, plus a consistency constraint to keep those trajectories coherent even with few steps. If the central claim holds, restoration becomes both quicker and more robust to degradations never seen in training. Readers would care because everyday image cleanup tasks like removing rain or noise could then run efficiently on devices while still preserving sharp details.

Core claim

IR-Flow constructs multilevel data distribution flows and cumulative velocity fields to learn transport trajectories that guide images from varying degradation levels toward clean targets, supported by a multi-step consistency constraint. The work shows that directly establishing a linear transport flow between degraded and clean image domains enables fast inference with only a few sampling steps while also improving adaptability to out-of-distribution degradations, as demonstrated through competitive results on deraining, denoising, and raindrop removal.

What carries the argument

Multilevel data distribution flows paired with cumulative velocity fields that define linear transport trajectories across degradation levels.

If this is right

  • Restoration achieves competitive quantitative scores using only a few sampling steps instead of many iterative ones.
  • Performance holds up better on degradations outside the training distribution than prior discriminative or generative baselines.
  • The approach maintains a strong balance between pixel-level accuracy and perceptual quality across tested tasks.
  • A single framework handles multiple restoration problems like deraining, denoising, and raindrop removal without task-specific redesign.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same linear transport idea could be tested on related inverse problems such as deblurring or super-resolution to check if few-step inference generalizes.
  • The consistency constraint might reduce variance in other generative sampling pipelines that currently need many steps.
  • Lower step counts open the possibility of running high-quality restoration directly on mobile hardware for live photo editing.

Load-bearing premise

That flows built across multiple degradation levels can steer partially restored images to the final clean version without creating artifacts or losing details when the degradation type is new.

What would settle it

A controlled test on an unseen degradation combination, such as rain plus sensor noise, where the outputs show visibly more artifacts or detail loss than a standard single-step discriminative network produces on the same inputs.

Figures

Figures reproduced from arXiv: 2604.19680 by Dong Li, Jie Huang, Jie Xiao, Xin Lu, Xueyang Fu, Zihao Fan.

Figure 1
Figure 1. Figure 1: Comparison of discriminative, SDE-based and our IR [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall framework of IR-Flow. The approach achieves image restoration through Rectified Flow, refined by cumulative velocity [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of two velocity fields. Standard velocity fields represents the instantaneous tangent directions. Our representation [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visual results of IR-Flow and competing approaches. The inference steps are set to 2 for deraining and raindrop removal, and 4 for [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visual comparisons between standard velocity and our [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Training and Performance curves of the IR-Flow on [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visual results of our IR-Flow method of dehazing on the RESIDE-6k [ [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visual results of the model trained on Rain100H evaluated on the Rain100L and real-world Internet dataset for out-of-distribution [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Multi-step solution results on denoising. As the number of solving steps increases, the details and textures of the recovery results [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Multi-step solution results on denoising. As the number of solving steps increases, the details and textures of the recovery results [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Visual results of the model on the real world deblurring GoPro dataset. [PITH_FULL_IMAGE:figures/full_fig_p019_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: More visual results on Rainy dataset. 6 [PITH_FULL_IMAGE:figures/full_fig_p019_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: More visual results on Noisy dataset. GT LQ RDDM Resfusion IR-Flow(ours) [PITH_FULL_IMAGE:figures/full_fig_p020_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Qualitative comparison results on Raindrop dataset. [PITH_FULL_IMAGE:figures/full_fig_p020_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: More visual results between 1-Rectified Flow standard velocity and our velocity on CIFAR10 (32 × 32) dataset. [PITH_FULL_IMAGE:figures/full_fig_p021_15.png] view at source ↗
read the original abstract

In image restoration, single-step discriminative mappings often lack fine details via expectation learning, whereas generative paradigms suffer from inefficient multi-step sampling and noise-residual coupling. To address this dilemma, we propose IR-Flow, a novel image restoration method based on Rectified Flow that serves as a unified framework bridging the gap between discriminative and generative paradigms. Specifically, we first construct multilevel data distribution flows, which expand the ability of models to learn from and adapt to various levels of degradation. Subsequently, cumulative velocity fields are proposed to learn transport trajectories across varying degradation levels, guiding intermediate states toward the clean target, while a multi-step consistency constraint is presented to enforce trajectory coherence and boost few-step restoration performance. We show that directly establishing a linear transport flow between degraded and clean image domains not only enables fast inference but also improves adaptability to out-of-distribution degradations. Extensive evaluations on deraining, denoising and raindrop removal tasks demonstrate that IR-Flow achieves competitive quantitative results with only a few sampling steps, offering an efficient and flexible framework that maintains an excellent distortion-perception balance. Our code is available at https://github.com/fanzh03/IR-Flow.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript proposes IR-Flow, a rectified-flow framework for image restoration that constructs multilevel data distribution flows to handle varying degradation levels, defines cumulative velocity fields to learn transport trajectories from degraded to clean domains, and adds a multi-step consistency constraint to improve few-step sampling. It claims this linear transport approach bridges discriminative and generative paradigms, enables fast inference, and improves adaptability to out-of-distribution degradations, while delivering competitive quantitative results on deraining, denoising, and raindrop removal tasks.

Significance. If the proposed components are shown to deliver the claimed benefits without introducing artifacts or detail loss, the work would provide a practical unified framework that combines the efficiency of few-step sampling with improved robustness, addressing a key tension in current image restoration methods. The open-sourced code is a positive factor for reproducibility.

major comments (2)
  1. [Abstract] Abstract: The load-bearing claim that 'directly establishing a linear transport flow ... improves adaptability to out-of-distribution degradations' is not supported by any cited ablation, OOD-specific test set, or error analysis; the construction of cumulative velocity fields (trained on specific degradation levels) does not automatically guarantee reliable extrapolation to unseen degradations such as combined noise or non-linear rain, as noted in the stress-test concern.
  2. [Method] Method description: The multilevel data distribution flows and cumulative velocity fields are presented as guiding intermediate states to the clean target, but without explicit equations or a derivation showing how the multi-step consistency constraint prevents deviation under distribution shift, it is unclear whether the trajectories remain artifact-free for OOD cases.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and have prepared revisions to strengthen the presentation of our claims and clarify the technical details.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The load-bearing claim that 'directly establishing a linear transport flow ... improves adaptability to out-of-distribution degradations' is not supported by any cited ablation, OOD-specific test set, or error analysis; the construction of cumulative velocity fields (trained on specific degradation levels) does not automatically guarantee reliable extrapolation to unseen degradations such as combined noise or non-linear rain, as noted in the stress-test concern.

    Authors: We agree that the abstract claim would be more robust with explicit supporting evidence. While our main experiments across deraining, denoising, and raindrop removal already cover a range of degradation intensities, we acknowledge the absence of dedicated OOD test sets and error analysis in the original submission. In the revised manuscript we have added a new subsection on out-of-distribution evaluation, including combined noise-plus-rain and non-linear rain patterns. These results show that IR-Flow retains higher PSNR/SSIM and fewer perceptual artifacts than competing methods, consistent with the linear-transport hypothesis. We have also inserted a brief error-analysis paragraph and softened the abstract wording to “suggests improved adaptability to out-of-distribution degradations” to reflect the new empirical support. revision: yes

  2. Referee: [Method] Method description: The multilevel data distribution flows and cumulative velocity fields are presented as guiding intermediate states to the clean target, but without explicit equations or a derivation showing how the multi-step consistency constraint prevents deviation under distribution shift, it is unclear whether the trajectories remain artifact-free for OOD cases.

    Authors: We accept that the method section would benefit from greater mathematical precision. In the revision we have inserted the missing explicit equations: the multilevel flow construction is now written as a family of linear interpolants parameterized by degradation level; the cumulative velocity field is defined as the integral of the learned velocity over these levels; and the multi-step consistency constraint is expressed as an L2 penalty between the single-step and multi-step trajectory endpoints. We further provide a short derivation in the main text (with full proof in the appendix) showing that the constraint bounds the deviation of the predicted velocity under bounded distribution shift, thereby helping keep trajectories artifact-free. These additions directly address the concern about OOD behavior. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation self-contained from external rectified-flow base

full rationale

The paper's chain begins with an external rectified-flow transport idea, then defines multilevel data distribution flows, cumulative velocity fields, and a multi-step consistency constraint as new architectural components. These are trained and evaluated on restoration tasks; the claims of fast inference and OOD adaptability are presented as empirical results of the construction rather than tautological redefinitions or fitted inputs renamed as predictions. No load-bearing self-citations, uniqueness theorems, or ansatzes imported from the authors' prior work appear in the provided text. The method remains falsifiable via standard benchmarks and does not reduce any key equation to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review prevents identification of concrete free parameters or axioms; the method appears to rest on the standard rectified-flow transport assumption plus new but unquantified components.

pith-pipeline@v0.9.0 · 5514 in / 1065 out tokens · 31962 ms · 2026-05-10T02:42:40.933620+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

87 extracted references · 87 canonical work pages

  1. [1]

    Abdelrahman Abdelhamed, Stephen Lin, and Michael S. Brown. A high-quality denoising dataset for smartphone cameras. In2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1692–1700, 2018. 7, 8, 2

    work page 2018

  2. [2]

    Ntire 2017 challenge on single image super-resolution: Dataset and study

    Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In2017 IEEE Conference on Computer Vision and Pattern Recogni- tion Workshops (CVPRW), pages 1122–1131, 2017. 6, 7, 8, 2

    work page 2017

  3. [3]

    Contour detection and hierarchical image segmen- tation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(5):898–916, 2011

    Pablo Arbel´aez, Michael Maire, Charless Fowlkes, and Jiten- dra Malik. Contour detection and hierarchical image segmen- tation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(5):898–916, 2011. 6

    work page 2011

  4. [4]

    The perception-distortion tradeoff

    Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6228–6237, 2018. 4

    work page 2018

  5. [5]

    Pre-trained image processing transformer

    Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12294–12305, 2021. 3

    work page 2021

  6. [6]

    Simple baselines for image restoration

    Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. InComputer Vision – ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part VII, page 17–33, Berlin, Heidelberg, 2022. Springer-Verlag. 3, 2

    work page 2022

  7. [7]

    A comparative study of image restoration networks for general backbone network design

    Xiangyu Chen, Zheyuan Li, Yuandong Pu, Yihao Liu, Jiantao Zhou, Yu Qiao, and Chao Dong. A comparative study of image restoration networks for general backbone network design. InComputer Vision – ECCV 2024: 18th European Conference, Milan, Italy, September 29–October 4, 2024, Pro- ceedings, Part LXXI, page 74–91, Berlin, Heidelberg, 2024. Springer-Verlag. 3

    work page 2024

  8. [8]

    Inversion by direct iteration: An alternative to denoising diffusion for image restoration.Transactions on Machine Learning Re- search, 2023

    Mauricio Delbracio and Peyman Milanfar. Inversion by direct iteration: An alternative to denoising diffusion for image restoration.Transactions on Machine Learning Re- search, 2023. https://openreview.net/forum? id=VmyFF5lL3F. 3, 7, 1

    work page 2023

  9. [9]

    Diffusion mod- els beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion mod- els beat gans on image synthesis. InAdvances in Neural Information Processing Systems, pages 8780–8794. Curran Associates, Inc., 2021. 3

    work page 2021

  10. [10]

    Image super-resolution using deep convolutional net- works.IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(2):295–307, 2016

    Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang. Image super-resolution using deep convolutional net- works.IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(2):295–307, 2016. 1

    work page 2016

  11. [11]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim En- tezari, Jonas M¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. InForty- first International Conference on Machine Learning, 2024. 3

    work page 2024

  12. [12]

    Vivid: Video virtual try-on using diffusion models, 2024

    Zixun Fang, Wei Zhai, Aimin Su, Hongliang Song, Kai Zhu, Mao Wang, Yu Chen, Zhiheng Liu, Yang Cao, and Zheng-Jun Zha. Vivid: Video virtual try-on using diffusion models, 2024. 3

    work page 2024

  13. [13]

    Generative diffu- sion prior for unified image restoration and enhancement

    Ben Fei, Zhaoyang Lyu, Liang Pan, Junzhe Zhang, Weidong Yang, Tianyue Luo, Bo Zhang, and Bo Dai. Generative diffu- sion prior for unified image restoration and enhancement. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9935–9946, 2023. 3

    work page 2023

  14. [14]

    Kodak lossless true color image suite, 1999

    Rich Franzen. Kodak lossless true color image suite, 1999. 6, 9, 3

    work page 1999

  15. [15]

    Image restoration by denoising diffusion models with iteratively preconditioned guidance

    Tomer Garber and Tom Tirer. Image restoration by denoising diffusion models with iteratively preconditioned guidance. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25245–25254, 2024. 3

    work page 2024

  16. [16]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bern- hard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Proceedings of the 31st International Conference on Neural Information Processing Systems, page 6629–6640, 2017. 6

    work page 2017

  17. [17]

    Denoising diffu- sion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models. InAdvances in Neural Information Processing Systems, pages 6840–6851, 2020. 1, 3

    work page 2020

  18. [18]

    Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M. Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2343–2353, 2022. 3

    work page 2022

  19. [19]

    Denoising diffusion restoration models

    Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. InAdvances in Neural Information Processing Systems, pages 23593–23606,

    work page

  20. [20]

    Musiq: Multi-scale image quality transformer

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5128–5137, 2021. 7

    work page 2021

  21. [21]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 1, 2

    work page 2015

  22. [22]

    Learning multi- ple layers of features from tiny images

    Alex Krizhevsky. Learning multi- ple layers of features from tiny images. https://api.semanticscholar.org/CorpusID:18268744,

    work page

  23. [23]

    Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better

    Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. In2019 IEEE/CVF International Confer- ence on Computer Vision (ICCV), pages 8877–8886, 2019. 3, 2

    work page 2019

  24. [24]

    Srdiff: Single image super-resolution with diffusion probabilistic models

    Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47–59, 2022. 3

    work page 2022

  25. [25]

    Swinir: Image restoration using swin transformer

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 1833–1844, 2021. 1

    work page 2021

  26. [26]

    Swinir: Image restoration 10 using swin transformer

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration 10 using swin transformer. In2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 1833–1844, 2021. 3

    work page 2021

  27. [27]

    Enhanced deep residual networks for single image super-resolution

    Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. In2017 IEEE Conference on Com- puter Vision and Pattern Recognition Workshops (CVPRW), pages 1132–1140, 2017. 6

    work page 2017

  28. [28]

    Score- based generative modeling through stochastic evolution equa- tions in hilbert spaces

    Sungbin Lim, EUN BI YOON, Taehyun Byun, Taewon Kang, Seungwoo Kim, Kyungjae Lee, and Sungjoon Choi. Score- based generative modeling through stochastic evolution equa- tions in hilbert spaces. InAdvances in Neural Information Processing Systems, pages 37799–37812, 2023. 3

    work page 2023

  29. [29]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maxim- ilian Nickel, and Matthew Le. Flow matching for genera- tive modeling. InThe Eleventh International Conference on Learning Representations, 2023. 3

    work page 2023

  30. [30]

    On the classification-distortion-perception tradeoff

    Dong Liu, Haochen Zhang, and Zhiwei Xiong. On the classification-distortion-perception tradeoff. InAdvances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 4

    work page 2019

  31. [31]

    Theodorou, Weili Nie, and Anima Anandkumar

    Guan-Horng Liu, Arash Vahdat, De-An Huang, Evangelos A. Theodorou, Weili Nie, and Anima Anandkumar. I2sb: image- to-image schr ¨odinger bridge. InProceedings of the 40th International Conference on Machine Learning, Honolulu, Hawaii, USA, 2023. JMLR.org. 3, 1

    work page 2023

  32. [32]

    Residual denoising diffusion mod- els

    Jiawei Liu, Qiang Wang, Huijie Fan, Yinong Wang, Yandong Tang, and Liangqiong Qu. Residual denoising diffusion mod- els. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2773–2783, 2024. 2, 3, 6, 9, 1

    work page 2024

  33. [33]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, and qiang liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. InThe Eleventh International Conference on Learning Representations, 2023. 3, 6

    work page 2023

  34. [34]

    Dpm-solver++: Fast solver for guided sam- pling of diffusion probabilistic models.Machine Intelligence Research, 22(4):730–751, 2025

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sam- pling of diffusion probabilistic models.Machine Intelligence Research, 22(4):730–751, 2025. 3

    work page 2025

  35. [35]

    Gustafsson, Zheng Zhao, Jens Sj¨olund, and Thomas B

    Ziwei Luo, Fredrik K. Gustafsson, Zheng Zhao, Jens Sj¨olund, and Thomas B. Sch¨on. Image restoration with mean-reverting stochastic differential equations. InProceedings of the 40th International Conference on Machine Learning, Honolulu, Hawaii, USA, 2023. JMLR.org. 2, 3, 6, 7, 9, 1

    work page 2023

  36. [36]

    Gustafsson, Zheng Zhao, Jens Sj¨olund, and Thomas B

    Ziwei Luo, Fredrik K. Gustafsson, Zheng Zhao, Jens Sj¨olund, and Thomas B. Sch¨on. Controlling vision-language models for multi-task image restoration. InThe Twelfth International Conference on Learning Representations, 2024. 7, 9

    work page 2024

  37. [37]

    Waterloo exploration database: New challenges for image quality as- sessment models.IEEE Transactions on Image Processing, 26(2):1004–1016, 2017

    Kede Ma, Zhengfang Duanmu, Qingbo Wu, Zhou Wang, Hongwei Yong, Hongliang Li, and Lei Zhang. Waterloo exploration database: New challenges for image quality as- sessment models.IEEE Transactions on Image Processing, 26(2):1004–1016, 2017. 6

    work page 2017

  38. [38]

    Martin, C

    D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. InProceedings Eighth IEEE International Confer- ence on Computer Vision. ICCV 2001, pages 416–423 vol.2,

    work page 2001

  39. [39]

    Deep multi-scale convolutional neural network for dynamic scene deblurring

    Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 257–265, 2017. 3, 7, 2

    work page 2017

  40. [40]

    Posterior- mean rectified flow: Towards minimum MSE photo-realistic image restoration

    Guy Ohayon, Tomer Michaeli, and Michael Elad. Posterior- mean rectified flow: Towards minimum MSE photo-realistic image restoration. InThe Thirteenth International Conference on Learning Representations, 2025. 3

    work page 2025

  41. [41]

    Restoring vision in adverse weather conditions with patch-based denoising diffusion models.IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1–12, 2023

    Ozan ¨Ozdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models.IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1–12, 2023. 6

    work page 2023

  42. [42]

    Tan, Wenhan Yang, Jiajun Su, and Jiaying Liu

    Rui Qian, Robby T. Tan, Wenhan Yang, Jiajun Su, and Jiaying Liu. Attentive generative adversarial network for raindrop removal from a single image. In2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2482– 2491, 2018. 6

    work page 2018

  43. [43]

    Ffa-net: Feature fusion attention network for single image dehazing.Proceedings of the AAAI Conference on Artificial Intelligence, 34(07):11908–11915, 2020

    Xu Qin, Zhilin Wang, Yuanchao Bai, Xiaodong Xie, and Huizhu Jia. Ffa-net: Feature fusion attention network for single image dehazing.Proceedings of the AAAI Conference on Artificial Intelligence, 34(07):11908–11915, 2020. 7, 8, 2

    work page 2020

  44. [44]

    Deep learning for seeing through window with raindrops

    Yuhui Quan, Shijie Deng, Yixin Chen, and Hui Ji. Deep learning for seeing through window with raindrops. In2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2463–2471, 2019. 3

    work page 2019

  45. [45]

    Progressive image deraining networks: A better and simpler baseline

    Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng. Progressive image deraining networks: A better and simpler baseline. In2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3932–3941, 2019. 3, 6

    work page 2019

  46. [46]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. In2022 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR), pages 10674–10685, 2022. 3

    work page 2022

  47. [47]

    Fleet, and Mohammad Norouzi

    Chitwan Saharia, Jonathan Ho, William Chan, Tim Sali- mans, David J. Fleet, and Mohammad Norouzi. Image super- resolution via iterative refinement.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713–4726,

    work page

  48. [48]

    Resfusion: Denoising diffusion probabilistic models for im- age restoration based on prior residual noise

    Zhenning Shi, Haoshuai Zheng, Chen Xu, Changsheng Dong, Bin Pan, Xueshuo Xie, Along He, Tao Li, and Huazhu Fu. Resfusion: Denoising diffusion probabilistic models for im- age restoration based on prior residual noise. InAdvances in Neural Information Processing Systems, pages 130664– 130693, 2024. 2, 3, 6, 9, 1

    work page 2024

  49. [49]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. InInternational Conference on Learning Representations, 2021. 1, 3, 6, 7, 2

    work page 2021

  50. [50]

    Nima: Neural image assessment.IEEE Transactions on Image Processing, 27(8): 3998–4011, 2018

    Hossein Talebi and Peyman Milanfar. Nima: Neural image assessment.IEEE Transactions on Image Processing, 27(8): 3998–4011, 2018. 7

    work page 2018

  51. [51]

    Maxim: Multi- 11 axis mlp for image processing

    Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Pey- man Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi- 11 axis mlp for image processing. In2022 IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), pages 5759–5770, 2022. 1, 6

    work page 2022

  52. [52]

    Bingshu Wang, Ze Wang, Wenjie Liu, Xiaoshui Huang, C. L. Philip Chen, and Yue Zhao. Ddsr-net: Direct document shadow removal leveraging multi-scale attention.Machine Intelligence Research, 22(3):452–465, 2025. 3

    work page 2025

  53. [53]

    Zero-shot image restoration using denoising diffusion null-space model

    Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. In The Eleventh International Conference on Learning Repre- sentations, 2023. 3

    work page 2023

  54. [54]

    Bovik, H.R

    Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity.IEEE Transactions on Image Processing, 13(4): 600–612, 2004. 6

    work page 2004

  55. [55]

    Uformer: A gen- eral u-shaped transformer for image restoration

    Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A gen- eral u-shaped transformer for image restoration. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17662–17672, 2022. 3

    work page 2022

  56. [56]

    Dimakis, and Peyman Milanfar

    Jay Whang, Mauricio Delbracio, Hossein Talebi, Chitwan Saharia, Alexandros G. Dimakis, and Peyman Milanfar. De- blurring via stochastic refinement. In2022 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR), pages 16272–16282, 2022. 3

    work page 2022

  57. [57]

    Fully 1 × 1 convolutional network for lightweight image super- resolution.Machine Intelligence Research, 21(6):1062–1076,

    Gang Wu, Junjun Jiang, Kui Jiang, and Xianming Liu. Fully 1 × 1 convolutional network for lightweight image super- resolution.Machine Intelligence Research, 21(6):1062–1076,

    work page

  58. [58]

    Advanced his- togram equalization based on a hybrid saliency map and novel visual prior.Machine Intelligence Research, 21(6): 1178–1191, 2024

    Yuanbin Wu, Shengkui Dai, and Zhan Ma. Advanced his- togram equalization based on a hybrid saliency map and novel visual prior.Machine Intelligence Research, 21(6): 1178–1191, 2024. 3

    work page 2024

  59. [59]

    Diffir: Efficient diffusion model for image restoration

    Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xing- long Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In 2023 IEEE/CVF International Conference on Computer Vi- sion (ICCV), pages 13049–13059, 2023. 2

    work page 2023

  60. [60]

    Diffir: Efficient diffusion model for image restoration

    Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xing- long Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In 2023 IEEE/CVF International Conference on Computer Vi- sion (ICCV), pages 13049–13059, 2023. 3

    work page 2023

  61. [61]

    Image de-raining transformer.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):12978– 12995, 2023

    Jie Xiao, Xueyang Fu, Aiping Liu, Feng Wu, and Zheng-Jun Zha. Image de-raining transformer.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):12978– 12995, 2023. 3, 6

    work page 2023

  62. [62]

    Dreamclean: Restoring clean image using deep diffusion prior

    Jie Xiao, Ruili Feng, Han Zhang, Zhiheng Liu, Zhantao Yang, Yurui Zhu, Xueyang Fu, Kai Zhu, Yu Liu, and Zheng-Jun Zha. Dreamclean: Restoring clean image using deep diffusion prior. InThe Twelfth International Conference on Learning Representations, 2024. 3

    work page 2024

  63. [63]

    Dual frequency transformer for efficient sdr-to-hdr translation.Machine Intel- ligence Research, 21(3):538–548, 2024

    Gang Xu, Qibin Hou, and Ming-Ming Cheng. Dual frequency transformer for efficient sdr-to-hdr translation.Machine Intel- ligence Research, 21(3):538–548, 2024. 3

    work page 2024

  64. [64]

    Tan, Jiashi Feng, Jiaying Liu, Zong- ming Guo, and Shuicheng Yan

    Wenhan Yang, Robby T. Tan, Jiashi Feng, Jiaying Liu, Zong- ming Guo, and Shuicheng Yan. Deep joint rain detection and removal from a single image. In2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1685–1694, 2017. 6, 9

    work page 2017

  65. [65]

    Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu

    Wenhan Yang, Robby T. Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu. Joint rain detection and removal from a single image with contextualized deep net- works.IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(6):1377–1393, 2020. 3, 6

    work page 2020

  66. [66]

    Image restoration through generalized ornstein-uhlenbeck bridge

    Conghan Yue, Zhengwei Peng, Junlong Ma, Shiyan Du, Pengxu Wei, and Dongyu Zhang. Image restoration through generalized ornstein-uhlenbeck bridge. InProceedings of the 41st International Conference on Machine Learning, 2024. 3, 7

    work page 2024

  67. [67]

    Resshift: Efficient diffusion model for image super-resolution by resid- ual shifting

    Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Resshift: Efficient diffusion model for image super-resolution by resid- ual shifting. InAdvances in Neural Information Processing Systems, pages 13294–13307, 2023. 2, 3, 1

    work page 2023

  68. [68]

    Multi-stage progressive image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14816–14826, 2021. 3, 6

    work page 2021

  69. [69]

    Restormer: Efficient transformer for high-resolution image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, and Ming–Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5718–5729,

    work page

  70. [70]

    Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising.IEEE Transactions on Image Processing, 26(7):3142–3155, 2017

    Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising.IEEE Transactions on Image Processing, 26(7):3142–3155, 2017. 3, 6

    work page 2017

  71. [71]

    Ffdnet: Toward a fast and flexible solution for cnn-based image denoising

    Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608–4622,

    work page

  72. [72]

    Practical blind image denoising via swin-conv- unet and data synthesis.Machine Intelligence Research, 20 (6):822–836, 2023

    Kai Zhang, Yawei Li, Jingyun Liang, Jiezhang Cao, Yu- lun Zhang, Hao Tang, Deng-Ping Fan, Radu Timofte, and Luc Van Gool. Practical blind image denoising via swin-conv- unet and data synthesis.Machine Intelligence Research, 20 (6):822–836, 2023. 3

    work page 2023

  73. [73]

    Color demosaicking by local directional interpolation and nonlocal adaptive thresholding.Journal of Electronic Imaging, 20(2),

    Lei Zhang, Xiaolin Wu, Antoni Buades, and Xin Li. Color demosaicking by local directional interpolation and nonlocal adaptive thresholding.Journal of Electronic Imaging, 20(2),

    work page

  74. [74]

    Efros, Eli Shechtman, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep fea- tures as a perceptual metric. In2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586–595,

    work page

  75. [75]

    Blind image quality assessment via vision- language correspondence: A multitask learning perspective

    Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision- language correspondence: A multitask learning perspective. In2023 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 14071–14081, 2023. 7

    work page 2023

  76. [76]

    UniDB: A unified diffusion bridge 12 framework via stochastic optimal control

    Kaizhen Zhu, Mokai Pan, Yuexin Ma, Yanwei Fu, Jingyi Yu, Jingya Wang, and Ye Shi. UniDB: A unified diffusion bridge 12 framework via stochastic optimal control. InForty-second International Conference on Machine Learning, 2025. 3, 7

    work page 2025

  77. [77]

    Denoising diffusion models for plug-and-play image restoration

    Yuanzhi Zhu, Kai Zhang, Jingyun Liang, Jiezhang Cao, Bi- han Wen, Radu Timofte, and Luc Van Gool. Denoising diffusion models for plug-and-play image restoration. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1219–1229, 2023. 3

    work page 2023

  78. [78]

    Flowie: Efficient image enhancement via rectified flow

    Yixuan Zhu, Wenliang Zhao, Ao Li, Yansong Tang, Jie Zhou, and Jiwen Lu. Flowie: Efficient image enhancement via rectified flow. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13–22, 2024. 3 13

    work page 2024

  79. [79]

    The CNN-baseline uses the same network but directly input the low-quality image and output the high- quality image

    Implementation details We chose the same U-Net backbone network as used in IR- SDE [35]. The CNN-baseline uses the same network but directly input the low-quality image and output the high- quality image. For most tasks, we set the training patch- size to be 256×256 and use a batch size of 8. We used the Adam [ 21] optimizer with parameters β1 = 0.9 and β...

    work page

  80. [80]

    Unified Formulation of SDE-based Methods To unify the representation of various SDE-based generative formulations for restoration, we introduce a general parame- terization of the noisy latent variablex t as: xt =α t x0 +β t x1 +γ t ϵ,(13) where x0 is the clean data, x1 is the observed degraded image, ϵ∼ N(0, I) is Gaussian noise, and αt, βt, γt are time-...

    work page

Showing first 80 references.