Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Post-training a restoration model on the 30% hardest samples with a composite reward lifts accuracy by up to 3.43 dB on unseen degradations.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 19:20 UTC pith:YPKS2S7Z

load-bearing objection Plausible but under-identified: GRPO's contribution is not isolated from hard-sample L1 fine-tuning; the paper deserves a serious review but needs a key control. the 4 major comments →

arxiv 2512.00814 v3 pith:YPKS2S7Z submitted 2025-11-30 cs.CV

IRPO: Boosting Image Restoration via Post-training GRPO

Haoxuan Xu , Yi Liu , Tianfu Li , Ruolin Shen , Boyuan Jiang , Jinlong Peng , Donghao Luo , Xiaobin Hu
show 2 more authors
Shuicheng Yan Haoang Li
This is my paper
classification cs.CV
keywords image restorationpost-trainingreinforcement learningGRPOdata curationperceptual rewardall-in-one restorationout-of-domain generalization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper claims that image-restoration models, which are usually trained by pixel-wise fitting, can be substantially improved by a post-training stage adapted from reinforcement learning. It proposes selecting only the 30% worst-performing training samples and optimizing a stochastic version of the model with group-relative policy optimization, using a reward made of generic quality metrics, a vision-language expert judge, and task-specific heuristics. On standard benchmarks this adds roughly 0.83 dB average PSNR over the strong AdaIR baseline; on out-of-domain real-world datasets the gain is 3.43 dB. A sympathetic reader cares because it suggests that hard-sample selection plus non-differentiable perceptual rewards is a general recipe for making restoration models both more accurate and more generalizable.

Core claim

The authors claim that a deterministic image-restoration network can be repurposed as a stochastic policy and improved through GRPO-based post-training. They identify two pillars: data-wise, training only on the 30% underperforming samples from the pre-training stage yields both higher accuracy and threefold faster training than full-data fine-tuning; reward-wise, a composite reward combining structural-fidelity metrics, a vision-language model's perceptual judgment, and task-specific low-level heuristics (gradient consistency for denoising, anisotropy for deraining, contrast for dehazing, sharpness for deblurring, exposure/color for low-light) provides a learnable, non-differentiable object

What carries the argument

The central mechanism is converting a deterministic all-in-one restoration backbone into a stochastic policy by adding lightweight policy heads that sample Beta-distributed controls (frequency-mask height/width and feature-fusion weights) for each restoration module. These actions are optimized with GRPO, which forms a group-normalized relative advantage from several stochastic rollouts and applies a clipped policy-gradient objective plus KL regularization. The second pillar is data curation: selecting the 30% hardest samples, measured by the pre-trained model's restoration quality, so that the post-training signal concentrates where the model fails.

Load-bearing premise

The choice of 30% underperforming data and the reward weights were tuned using the same in-domain benchmarks that later report the gains, with no separate held-out validation split, so the recipe may be overfit to these tasks.

What would settle it

Apply the exact 30% hard-sample rule and the same reward weights to a new restoration task, such as single-image super-resolution, and compare against random-subset post-training on a held-out benchmark. If the hard-sample advantage disappears or out-of-domain gains reverse, the claimed principle is not general.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the claim holds, restoration models can be boosted without redesigning their architecture or training data pipeline—just by post-training on a curated hard subset with a composite reward.
  • Non-differentiable perceptual rewards can be used directly to optimize restoration quality, bypassing the need for differentiable surrogates of human preference.
  • The data-curation principle implies that post-training efficiency can improve severalfold by discarding easy samples, which may transfer to other low-level vision tasks.
  • The large out-of-domain gains suggest that reward-based post-training can reduce overfitting to synthetic training distributions and improve robustness to real-world degradations.

Where Pith is reading between the lines

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

  • A natural extension is to test the same 30% hard-sample rule and reward trio on other restoration tasks such as super-resolution or demosaicing; if the gains persist, the recipe is general, if not, the principle is task-specific.
  • The composite reward could be made even more powerful by replacing the fixed vision-language judge with a learned image-quality-assessment model that is itself updated during training, reducing reward-hacking risk.
  • The observed OOD improvement suggests that the policy learns a more transferable representation of degradation rather than memorizing training-domain statistics, which could be verified by probing the learned feature-fusion weights.
  • A practical consequence the authors do not draw explicitly: the same post-training framework might be applied to video restoration or blind restoration where pixel-wise losses are particularly prone to over-smoothing.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes IRPO, a GRPO-based post-training paradigm for image restoration. It re-frames post-training around two pillars: a data-oriented pillar that selects the 30% worst-performing training samples according to the pre-trained model, and a reward-oriented pillar that optimizes a composite reward (general perceptual/fidelity metrics, an Qwen-VL expert score, and task-specific heuristic rewards) via GRPO over a stochastic policy head attached to the AdaIR backbone. The final objective (Eq. 7) jointly optimizes the GRPO surrogate with an L1 supervised loss on the best-scoring rollout and a consistency loss. Experiments report gains over AdaIR on six in-domain benchmarks (average +0.83 dB PSNR in the five-degradation all-in-one setting, Table 2) and five OOD real-world benchmarks (average +3.43 dB, Table 3), with ablations on data fraction, reward components, and group size.

Significance. If the reported gains are robust, the paper offers a generally applicable post-training recipe for low-level vision that improves both in-domain accuracy and OOD generalization, and it provides a useful decomposition into data curation and reward design. The paper is commendably thorough in its experimental scope (single-task, three-degradation, five-degradation all-in-one, OOD), and it ships code and detailed hyperparameters. The OOD improvements, in particular, are striking and cannot be explained by simple metric overfitting alone. However, the causal role of the GRPO/reward machinery is underdetermined by the current ablations, and several evaluation choices risk overfitting to the reported benchmarks.

major comments (4)
  1. [Sec. 3.4, Eq. (7); Sec. 3.2, Eqs. (1)-(2)] The central claim that GRPO post-training is what drives the improvement is not isolated. The final objective jointly optimizes L_RL, L_sup, and L_cons on the hard-data subset. L_sup is an L1 loss to the rollout that scores best under a reward that includes PSNR and SSIM (Eq. (22), Appendix B.1). Thus a plausible alternative explanation is that hard-sample L1 fine-tuning alone, without any policy-gradient or reward machinery, produces most of the gain. The paper needs a control that removes L_RL while keeping L_sup/L_cons on D_hard, and a control that removes L_sup/L_cons while keeping L_RL. Table 4 ablates reward components but all rows appear to retain the supervised/consistency terms, so it cannot establish that the reward signal is the operative factor. This is load-bearing for the paper's central claim.
  2. [Sec. 4.4, Figure 5; Sec. 4.2, Table 2] The data-fraction hyperparameter (30%) and the group size G are selected by evaluating on the same all-in-one 5-task benchmark that is later used to report final results in Table 2. No held-out validation split is used, and the paper reports no run-to-run variance or error bars. Consequently, the '30% underperforming data' principle and the claim of a favorable efficiency-accuracy trade-off may be overfit to these specific benchmarks. The authors should either use a validation split for model selection or report multiple seeds with standard deviations on at least the main in-domain and OOD tables.
  3. [Sec. 3.3, Eq. (22); Appendix B.1] The reward includes PSNR and SSIM as components of R_gen, and the paper's primary reported metrics are PSNR/SSIM. Thus the in-domain gains on these metrics are partially optimized-for rather than independent predictions. The paper should explicitly acknowledge this circularity and provide at least one evaluation metric (e.g., LPIPS, or a no-reference quality score) that is not present in the reward, to show that the improvement is not merely a reward-following artifact. The OOD results partly mitigate this concern, but the same issue applies to the degree that PSNR/SSIM are in the reward.
  4. [Abstract vs. Sec. 4.2, Table 2; Table 3 caption vs. Sec. 4.3] The abstract states a 0.93 dB in-domain improvement, while the body and Table 2 report 0.83 dB. Similarly, the Table 3 caption says 'an average of 2.53 dB PSNR improvement,' but the text and the table numbers give 3.43 dB. These inconsistencies in the headline numbers need correction; if the 0.93 dB figure comes from a different configuration, that must be clarified.
minor comments (5)
  1. [Throughout] Please standardize the spelling of 'AdaIR' (the text and tables sometimes use 'Adair'), and fix the repeated typo in the code-availability statement ('Our code can be shown in...').
  2. [Sec. 4.4, Figure 5] Figure 5's right y-axis (post-training time) is described in GPU days, but the axis label appears as a raw number; please add units. Also, the figure legend would benefit from explicitly marking which line corresponds to random vs. underperforming subsets.
  3. [Sec. 3.3, Eq. (6)] The entropy bonus term is written as -τ H(π_θ), but the sign convention is not stated. Since L_RL is minimized (Eq. 7 uses L_total), please clarify whether this is a negative entropy bonus or a typo, and ensure the sign matches the implementation.
  4. [Appendix B.1] The PSNR mapping to [0,1] via fixed thresholds (τ_min, τ_max) is not specified. Provide the exact thresholds so the reward is reproducible.
  5. [Sec. 4.1, Implementation Details] The sentence 'λsup: star from 0.35 to 0.1 end' contains a typo ('star' for 'start'). Also specify the exact annealing schedule for λcons and whether it is linear per epoch or per step.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claims rest on several hand-chosen hyperparameters (data ratio, group size, reward weights, annealing schedules) and on two domain assumptions about the reliability of the Qwen judge and the task-specific heuristics. No free physical constants or invented entities are introduced.

free parameters (6)
  • hard_data_ratio = 30%
    Chosen by sweeping the fraction on the in-domain evaluation benchmarks (Figure 5); no held-out validation.
  • group_size_G = 4
    Chosen by ablation on test benchmarks (Table 5); G=4 and G=5 differ by only 0.01 dB, so the choice is within noise.
  • reward_weights = lambda_gen=0.6, lambda_qwen=0.1, lambda_task=0.3
    Fixed balancing hyperparameters in Eq. (22); no sensitivity analysis.
  • lambda_sup_anneal = 0.35 -> 0.1
    Anneal schedule for supervised L1 loss in Eq. (7); chosen by hand.
  • lambda_cons_anneal = 0.2 -> 0.05
    Anneal schedule for consistency loss; chosen by hand.
  • reward_constants = PSNR thresholds, 0.5 exposure tolerance, 0.6 color tolerance, entropy 0.01
    Arbitrary normalizing constants in Appendix B that affect reward scale and balance.
axioms (5)
  • standard math GRPO clipped surrogate objective and Beta-policy closed-form log-probability (Eqs. 6, 23) are valid.
    Background from DeepSeekMath; used as the RL update rule.
  • domain assumption Qwen2.5-VL-7B's 1-5 score is a sufficiently reliable perceptual reward for restoration post-training.
    The paper itself notes 'potential hallucinations from Qwen' (Sec 4.4); if the judge is noisy or biased, the R_qwen component distorts the policy.
  • domain assumption Task-specific heuristics (gradient consistency, anisotropy, contrast/sharpness closeness, exposure/color) are valid proxies for restoration quality.
    These heuristics are simple statistics of the restored image and GT; they are not shown to correlate with human perception beyond PSNR/SSIM.
  • domain assumption The worst-30% subset by initial PSNR is the optimal post-training data.
    This is the paper's data-oriented principle; its optimality is only demonstrated on the evaluation benchmarks, not on a held-out split.
  • domain assumption Adding Beta-distributed policy heads to AdaIR's AFLB modules preserves the pre-trained representation quality.
    The paper initializes from AdaIR and only fine-tunes with a small learning rate; no analysis of representation drift is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 22431 in / 13272 out tokens · 120439 ms · 2026-08-03T19:20:02.674319+00:00 · methodology

0 comments
read the original abstract

Post-training has become effective for high-level generation, but its role in low-level vision remains underexplored. Existing image restoration methods often rely on fixed pixel-wise fitting to ground-truth images, which can lead to over-smoothing and weak generalization. We propose IRPO, a GRPO-based post-training framework for deterministic restoration models. IRPO is built around two axes: data formulation and reward modeling. For data formulation, we select the 30% underperforming samples from the pre-training stage, which improves both accuracy and training efficiency. For reward modeling, we combine fidelity-oriented and quality-aware feedback with three components: a General Reward for structural fidelity, an Expert Reward that uses a Vision-Language Model as a coarse visual-quality judge, and a Restoration Reward for task-specific low-level cues. Experiments on six in-domain and five out-of-domain (OOD) benchmarks show that IRPO improves the AdaIR baseline by 0.93 dB on in-domain tasks and 3.43 dB on OOD settings. Our code can be shown in https://github.com/HaoxuanXU1024/IRPO.

Figures

Figures reproduced from arXiv: 2512.00814 by Boyuan Jiang, Donghao Luo, Haoang Li, Haoxuan Xu, Jinlong Peng, Ruolin Shen, Shuicheng Yan, Tianfu Li, Xiaobin Hu, Yi Liu.

Figure 1
Figure 1. Figure 1: An overview of our IRPO post-training paradigm and its [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The overview of our proposed post-training paradigm, visually structured around its two pillars. Pillar 1 (Data-Oriented, left): [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visual comparisons for different restoration tasks. The first row is derain, the second row is dehaze, and the third row is denoise [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visual comparisons on real-world datasets. From top to [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Post-training on underperforming vs. random subsets for the All-in-One five tasks. The left y-axis shows average PSNR (dB) and the right y-axis shows post-training time (GPU days); the dashed line marks the AdaIR baseline (30.2 dB). Balancing accuracy and cost, the 30% underperforming subset offers the best trade-off (31.4 dB in 4.2 days, comparable to 100% data but three times faster). 0k 20k 40k 60k 80k … view at source ↗
Figure 6
Figure 6. Figure 6: Training curves for our IRPO post-training ( [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Supplementary visual comparisons for image dehazing on synthetic datasets. Our method (IRPO) achieves the highest PSNR [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Supplementary visual comparisons for image denoising (noise level 50) on synthetic datasets. Our method (IRPO) consistently [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Supplementary visual comparisons for image deraining on synthetic datasets. IRPO achieves superior restoration quality. [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Bridging Information Asymmetry: A Hierarchical Framework for Deterministic Blind Face Restoration

    cs.CV 2026-01 conditional novelty 5.0

    Pref-Restore combines AR semantic tokens, a diffusion generator, and DiffusionNFT-style RL to make blind face restoration more consistent, but its deterministic-identity claim is weakened by self-referential rewards a...

Reference graph

Works this paper leans on

91 extracted references · 13 linked inside Pith · cited by 1 Pith paper

  1. [1]

    A high-quality denoising dataset for smartphone cameras

    Abdelrahman Abdelhamed, Stephen Lin, and Michael S Brown. A high-quality denoising dataset for smartphone cameras. InProceedings of the IEEE conference on com- puter vision and pattern recognition, pages 1692–1700,

  2. [2]

    Contour detection and hierarchical image seg- mentation.TPAMI, 2010

    Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Ji- tendra Malik. Contour detection and hierarchical image seg- mentation.TPAMI, 2010. 5

    2010

  3. [3]

    Not just streaks: Towards ground truth for single image derain- ing

    Yunhao Ba, Howard Zhang, Ethan Yang, Akira Suzuki, Arnold Pfahnl, Chethan Chinder Chandrappa, Celso M De Melo, Suya You, Stefano Soatto, Alex Wong, et al. Not just streaks: Towards ground truth for single image derain- ing. InEuropean Conference on Computer Vision, pages 723–740. Springer, 2022. 6

    2022

  4. [4]

    Training a helpful and harmless assistant with reinforcement learning from human feedback.arXiv preprint arXiv:2204.05862, 2022

    Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback.arXiv preprint arXiv:2204.05862, 2022. 3

    Pith/arXiv arXiv 2022

  5. [5]

    The perception-distortion tradeoff

    Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 6228–6237, 2018. 2, 3

    2018

  6. [6]

    Dehazenet: An end-to-end system for single image haze removal.TIP, 2016

    Bolun Cai, Xiangmin Xu, Kui Jia, Chunmei Qing, and Dacheng Tao. Dehazenet: An end-to-end system for single image haze removal.TIP, 2016. 5

    2016

  7. [7]

    Hinet: Half instance normalization network for image restoration

    Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Cheng- peng Chen. Hinet: Half instance normalization network for image restoration. InCVPR Workshops, 2021. 5

    2021

  8. [8]

    Simple baselines for image restoration

    Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. InECCV, 2022. 5

    2022

  9. [9]

    Liang Chen, Lei Li, Haozhe Zhao, and Yifan Song. Vinci. r1-v: Reinforcing super generalization ability in vision- language models with less than $3, 2025. 3

    2025

  10. [10]

    AdaIR: Adap- tive all-in-one image restoration via frequency mining and modulation

    Yuning Cui, Syed Waqas Zamir, Salman Khan, Alois Knoll, Mubarak Shah, and Fahad Shahbaz Khan. AdaIR: Adap- tive all-in-one image restoration via frequency mining and modulation. InThe Thirteenth International Conference on Learning Representations, 2025. 2, 4, 5, 6, 3

    2025

  11. [11]

    Color image denoising via sparse 3d col- laborative filtering with grouping constraint in luminance- chrominance space

    Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Color image denoising via sparse 3d col- laborative filtering with grouping constraint in luminance- chrominance space. InICIP, 2007. 5

    2007

  12. [12]

    Image super-resolution using deep convolutional net- works.IEEE transactions on pattern analysis and machine intelligence, 38(2):295–307, 2015

    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, 2015. 2

    2015

  13. [13]

    Fd-gan: Generative adversarial networks with fusion- discriminator for single image dehazing

    Yu Dong, Yihao Liu, He Zhang, Shifeng Chen, and Yu Qiao. Fd-gan: Generative adversarial networks with fusion- discriminator for single image dehazing. InAAAI, 2020. 5, 4

    2020

  14. [14]

    A general decoupled learn- ing framework for parameterized image operators.TPAMI,

    Qingnan Fan, Dongdong Chen, Lu Yuan, Gang Hua, Neng- hai Yu, and Baoquan Chen. A general decoupled learn- ing framework for parameterized image operators.TPAMI,

  15. [15]

    Dy- namic scene deblurring with parameter selective sharing and nested skip connections

    Hongyun Gao, Xin Tao, Xiaoyong Shen, and Jiaya Jia. Dy- namic scene deblurring with parameter selective sharing and nested skip connections. InCVPR, 2019. 5, 4

    2019

  16. [16]

    Search, verify and feedback: Towards next generation post- training paradigm of foundation models via verifier engineer- ing.arXiv preprint arXiv:2411.11504, 2024

    Xinyan Guan, Yanjiang Liu, Xinyu Lu, Boxi Cao, Ben He, Xianpei Han, Le Sun, Jie Lou, Bowen Yu, Yaojie Lu, et al. Search, verify and feedback: Towards next generation post- training paradigm of foundation models via verifier engineer- ing.arXiv preprint arXiv:2411.11504, 2024. 1

    Pith/arXiv arXiv 2024

  17. [17]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025. 3

    Pith/arXiv arXiv 2025

  18. [18]

    Toward convolutional blind denoising of real pho- tographs

    Shi Guo, Zifei Yan, Kai Zhang, Wangmeng Zuo, and Lei Zhang. Toward convolutional blind denoising of real pho- tographs. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1712–1722,

  19. [19]

    Face super-resolution guided by 3d facial priors

    Xiaobin Hu, Wenqi Ren, John LaMaster, Xiaochun Cao, Xi- aoming Li, Zechao Li, Bjoern Menze, and Wei Liu. Face super-resolution guided by 3d facial priors. InEuropean Conference on Computer Vision, pages 763–780. Springer,

  20. [20]

    Face restoration via plug-and-play 3d facial priors.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):8910–8926, 2021

    Xiaobin Hu, Wenqi Ren, Jiaolong Yang, Xiaochun Cao, David Wipf, Bjoern Menze, Xin Tong, and Hongbin Zha. Face restoration via plug-and-play 3d facial priors.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):8910–8926, 2021. 3

    2021

  21. [21]

    Single image super-resolution from transformed self-exemplars

    Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, 2015. 5

    2015

  22. [22]

    Hunyuan3d-omni: A unified framework for controllable generation of 3d assets

    Team Hunyuan3D, Bowen Zhang, Chunchao Guo, Haolin Liu, Hongyu Yan, Huiwen Shi, Jingwei Huang, Junlin Yu, Kunhong Li, Penghao Wang, et al. Hunyuan3d-omni: A unified framework for controllable generation of 3d assets. arXiv preprint arXiv:2509.21245, 2025. 1

    arXiv 2025

  23. [23]

    Loli-street: Bench- marking low-light image enhancement and beyond

    Md Tanvir Islam, Inzamamul Alam, Simon S Woo, Saeed Anwar, IK Lee, and Khan Muhammad. Loli-street: Bench- marking low-light image enhancement and beyond. InPro- ceedings of the Asian Conference on Computer Vision, pages 1250–1267, 2024. 6

    2024

  24. [24]

    Openai o1 system card.arXiv preprint arXiv:2412.16720, 2024

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richard- son, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card.arXiv preprint arXiv:2412.16720, 2024. 3

    Pith/arXiv arXiv 2024

  25. [25]

    Supervised learning of image restoration with convolutional networks

    Viren Jain, Joseph F Murray, Fabian Roth, Srinivas Turaga, Valentin Zhigulin, Kevin L Briggman, Moritz N Helm- staedter, Winfried Denk, and H Sebastian Seung. Supervised learning of image restoration with convolutional networks. In2007 IEEE 11th International Conference on Computer Vision, pages 1–8. IEEE, 2007. 2

    2007

  26. [26]

    Sonic: Shifting focus to global au- dio perception in portrait animation

    Xiaozhong Ji, Xiaobin Hu, Zhihong Xu, Junwei Zhu, Chum- ing Lin, Qingdong He, Jiangning Zhang, Donghao Luo, Yi 9 Chen, Qin Lin, et al. Sonic: Shifting focus to global au- dio perception in portrait animation. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 193–203, 2025. 1

    2025

  27. [27]

    Multi-scale progressive fusion network for single image deraining

    Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Baojin Huang, Yimin Luo, Jiayi Ma, and Junjun Jiang. Multi-scale progressive fusion network for single image deraining. In CVPR, 2020. 5

    2020

  28. [28]

    Perceptual losses for real-time style transfer and super-resolution

    Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694–711. Springer, 2016. 2, 3

    2016

  29. [29]

    Can grpo boost complex mul- timodal table understanding? InProceedings of the 2025 Conference on Empirical Methods in Natural Language Pro- cessing, pages 12642–12655, 2025

    Xiaoqiang Kang, Shengen Wu, Zimu Wang, Yilin Liu, Xi- aobo Jin, Kaizhu Huang, Wei Wang, Yutao Yue, Xiaowei Huang, and Qiufeng Wang. Can grpo boost complex mul- timodal table understanding? InProceedings of the 2025 Conference on Empirical Methods in Natural Language Pro- cessing, pages 12642–12655, 2025. 3

    2025

  30. [30]

    A survey of post-training scaling in large language models

    Hanyu Lai, Xiao Liu, Junjie Gao, Jiale Cheng, Zehan Qi, Yi- fan Xu, Shuntian Yao, Dan Zhang, Jinhua Du, Zhenyu Hou, et al. A survey of post-training scaling in large language models. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2771–2791, 2025. 1

    2025

  31. [31]

    Noise2noise: Learning image restoration without clean data

    Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2noise: Learning image restoration without clean data. arXiv preprint arXiv:1803.04189, 2018. 2, 3

    Pith/arXiv arXiv 2018

  32. [32]

    Benchmarking single- image dehazing and beyond.TIP, 2018

    Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single- image dehazing and beyond.TIP, 2018. 5, 4

    2018

  33. [33]

    All-in-one image restoration for unknown cor- ruption

    Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown cor- ruption. InCVPR, 2022. 3, 5, 4

    2022

  34. [34]

    Real-world deep local motion deblur- ring

    Haoying Li, Ziran Zhang, Tingting Jiang, Peng Luo, Huajun Feng, and Zhihai Xu. Real-world deep local motion deblur- ring. Inproceedings of the AAAI conference on artificial intelligence, pages 1314–1322, 2023. 6

    2023

  35. [35]

    T2v- turbo-v2: Enhancing video generation model post-training through data, reward, and conditional guidance design.arXiv preprint arXiv:2410.05677, 2024

    Jiachen Li, Qian Long, Jian Zheng, Xiaofeng Gao, Robinson Piramuthu, Wenhu Chen, and William Yang Wang. T2v- turbo-v2: Enhancing video generation model post-training through data, reward, and conditional guidance design.arXiv preprint arXiv:2410.05677, 2024. 1, 3

    arXiv 2024

  36. [36]

    Swinir: Image restoration us- ing swin transformer

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration us- ing swin transformer. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 1833–1844,

  37. [37]

    Vrt: A video restoration transformer.IEEE Transactions on Image Processing, 33:2171–2182, 2024

    Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer.IEEE Transactions on Image Processing, 33:2171–2182, 2024. 3

    2024

  38. [38]

    Vton-handfit: Virtual try-on for arbi- trary hand pose guided by hand priors embedding

    Yujie Liang, Xiaobin Hu, Boyuan Jiang, Donghao Luo, Xu Peng, Kai Wu, Chengming Xu, Wenhui Han, Taisong Jin, Chengjie Wang, et al. Vton-handfit: Virtual try-on for arbi- trary hand pose guided by hand priors embedding. InPro- ceedings of the Computer Vision and Pattern Recognition Conference, pages 22616–22626, 2025. 1

    2025

  39. [39]

    Diffusion adversarial post-training for one-step video generation.arXiv preprint arXiv:2501.08316,

    Shanchuan Lin, Xin Xia, Yuxi Ren, Ceyuan Yang, Xuefeng Xiao, and Lu Jiang. Diffusion adversarial post-training for one-step video generation.arXiv preprint arXiv:2501.08316,

    arXiv

  40. [40]

    Autoregressive adversarial post-training for real-time inter- active video generation.arXiv preprint arXiv:2506.09350,

    Shanchuan Lin, Ceyuan Yang, Hao He, Jianwen Jiang, Yuxi Ren, Xin Xia, Yang Zhao, Xuefeng Xiao, and Lu Jiang. Autoregressive adversarial post-training for real-time inter- active video generation.arXiv preprint arXiv:2506.09350,

    arXiv

  41. [41]

    Tape: Task-agnostic prior embedding for image restoration

    Lin Liu, Lingxi Xie, Xiaopeng Zhang, Shanxin Yuan, Xi- angyu Chen, Wengang Zhou, Houqiang Li, and Qi Tian. Tape: Task-agnostic prior embedding for image restoration. InECCV, 2022. 5

    2022

  42. [42]

    Moa-vr: A mixture- of-agents system towards all-in-one video restoration.arXiv preprint arXiv:2510.08508, 2025

    Lu Liu, Chunlei Cai, Shaocheng Shen, Jianfeng Liang, Weimin Ouyang, Tianxiao Ye, Jian Mao, Huiyu Duan, Jiangchao Yao, Xiaoyun Zhang, et al. Moa-vr: A mixture- of-agents system towards all-in-one video restoration.arXiv preprint arXiv:2510.08508, 2025. 2, 3

    arXiv 2025

  43. [43]

    Two-stage mamba-based diffusion model for image restora- tion.Scientific Reports, 15(1):22265, 2025

    Lei Liu, Luan Ma, Shuai Wang, Jun Wang, and Silas N Melo. Two-stage mamba-based diffusion model for image restora- tion.Scientific Reports, 15(1):22265, 2025. 2

    2025

  44. [44]

    Post-training quantization for vision trans- former.Advances in Neural Information Processing Systems, 34:28092–28103, 2021

    Zhenhua Liu, Yunhe Wang, Kai Han, Wei Zhang, Siwei Ma, and Wen Gao. Post-training quantization for vision trans- former.Advances in Neural Information Processing Systems, 34:28092–28103, 2021. 1, 3

    2021

  45. [45]

    Waterloo ex- ploration database: New challenges for image quality assess- ment models.TIP, 2016

    Kede Ma, Zhengfang Duanmu, Qingbo Wu, Zhou Wang, Hongwei Yong, Hongliang Li, and Lei Zhang. Waterloo ex- ploration database: New challenges for image quality assess- ment models.TIP, 2016. 5

    2016

  46. [46]

    A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics

    David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. InICCV, 2001. 5, 4

    2001

  47. [47]

    Deep generalized unfolding networks for image restoration

    Chong Mou, Qian Wang, and Jian Zhang. Deep generalized unfolding networks for image restoration. InCVPR, 2022. 5

    2022

  48. [48]

    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. InCVPR, 2017. 5

    2017

  49. [49]

    Promptir: Prompting for all-in- one image restoration.NeurIPS, 2023

    Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Promptir: Prompting for all-in- one image restoration.NeurIPS, 2023. 2, 5, 3, 4

    2023

  50. [50]

    Vln-r1: Vision-language navigation via reinforcement fine-tuning.arXiv preprint arXiv:2506.17221,

    Zhangyang Qi, Zhixiong Zhang, Yizhou Yu, Jiaqi Wang, and Hengshuang Zhao. Vln-r1: Vision-language navigation via reinforcement fine-tuning.arXiv preprint arXiv:2506.17221,

    Pith/arXiv arXiv

  51. [51]

    En- hanced pix2pix dehazing network

    Yanyun Qu, Yizi Chen, Jingying Huang, and Yuan Xie. En- hanced pix2pix dehazing network. InCVPR, 2019. 5

    2019

  52. [52]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021. 1

    2021

  53. [53]

    Direct preference optimization: Your language model is secretly a reward model.Advances in neural information processing systems, 36:53728–53741, 2023

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christo- pher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model.Advances in neural information processing systems, 36:53728–53741, 2023. 3 10

    2023

  54. [54]

    Single image dehazing via multi- scale convolutional neural networks

    Wenqi Ren, Si Liu, Hua Zhang, Jinshan Pan, Xiaochun Cao, and Ming-Hsuan Yang. Single image dehazing via multi- scale convolutional neural networks. InECCV, 2016. 5

    2016

  55. [55]

    Laion-5b: An open large-scale dataset for training next generation image-text models.Advances in neural in- formation processing systems, 35:25278–25294, 2022

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Worts- man, et al. Laion-5b: An open large-scale dataset for training next generation image-text models.Advances in neural in- formation processing systems, 35:25278–25294, 2022. 1

    2022

  56. [56]

    Post-training quantization on diffusion models

    Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 1972–1981, 2023. 1

    1972

  57. [57]

    Deepseekmath: Pushing the limits of math- ematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of math- ematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024. 2, 3

    Pith/arXiv arXiv 2024

  58. [58]

    Fine-grained image quality assessment for per- ceptual image restoration.arXiv preprint arXiv:2508.14475,

    Xiangfei Sheng, Xiaofeng Pan, Zhichao Yang, Pengfei Chen, and Leida Li. Fine-grained image quality assessment for per- ceptual image restoration.arXiv preprint arXiv:2508.14475,

    arXiv

  59. [59]

    Image de- noising using deep cnn with batch renormalization.Neural Networks, 2020

    Chunwei Tian, Yong Xu, and Wangmeng Zuo. Image de- noising using deep cnn with batch renormalization.Neural Networks, 2020. 5, 4

    2020

  60. [60]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth´ee Lacroix, Baptiste Rozi`ere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3

    Pith/arXiv arXiv 2023

  61. [61]

    Transweather: Transformer-based restoration of im- ages degraded by adverse weather conditions

    Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of im- ages degraded by adverse weather conditions. InCVPR,

  62. [62]

    Apisr: Anime production inspired real-world anime super-resolution

    Boyang Wang, Fengyu Yang, Xihang Yu, Chao Zhang, and Hanbin Zhao. Apisr: Anime production inspired real-world anime super-resolution. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25574–25584, 2024. 2

    2024

  63. [63]

    Esrgan: En- hanced super-resolution generative adversarial networks

    Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: En- hanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 3

    2018

  64. [64]

    Deep retinex decomposition for low-light enhancement

    Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retinex decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 5

    Pith/arXiv arXiv 2018

  65. [65]

    Boosting all-in-one image restoration via self-improved priv- ilege learning.arXiv preprint arXiv:2505.24207, 2025

    Gang Wu, Junjun Jiang, Kui Jiang, and Xianming Liu. Boosting all-in-one image restoration via self-improved priv- ilege learning.arXiv preprint arXiv:2505.24207, 2025. 2

    Pith/arXiv arXiv 2025

  66. [66]

    Cross-domain car detection model with integrated convolu- tional block attention mechanism.Image and Vision Com- puting, 140:104834, 2023

    Haoxuan Xu, Songning Lai, Xianyang Li, and Yang Yang. Cross-domain car detection model with integrated convolu- tional block attention mechanism.Image and Vision Com- puting, 140:104834, 2023. 2

    2023

  67. [67]

    Joint rain detection and removal from a single image with contextualized deep net- works.TPAMI, 2019

    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.TPAMI, 2019. 5, 4

    2019

  68. [68]

    R1-onevision: Advancing gen- eralized multimodal reasoning through cross-modal formal- ization.arXiv preprint arXiv:2503.10615, 2025

    Yi Yang, Xiaoxuan He, Hongkun Pan, Xiyan Jiang, Yan Deng, Xingtao Yang, Haoyu Lu, Dacheng Yin, Fengyun Rao, Minfeng Zhu, et al. R1-onevision: Advancing gen- eralized multimodal reasoning through cross-modal formal- ization.arXiv preprint arXiv:2503.10615, 2025. 2, 3

    Pith/arXiv arXiv 2025

  69. [69]

    Uncertainty guided multi-scale residual learning-using a cycle spinning cnn for single image de-raining

    Rajeev Yasarla and Vishal M Patel. Uncertainty guided multi-scale residual learning-using a cycle spinning cnn for single image de-raining. InCVPR, 2019. 5

    2019

  70. [70]

    Vla-r1: Enhancing rea- soning in vision-language-action models.arXiv preprint arXiv:2510.01623, 2025

    Angen Ye, Zeyu Zhang, Boyuan Wang, Xiaofeng Wang, Dapeng Zhang, and Zheng Zhu. Vla-r1: Enhancing rea- soning in vision-language-action models.arXiv preprint arXiv:2510.01623, 2025. 3

    arXiv 2025

  71. [71]

    Fera: Frequency-energy con- strained routing for effective diffusion adaptation fine- tuning.arXiv preprint arXiv:2511.17979, 2025

    Bo Yin, Xiaobin Hu, Xingyu Zhou, Peng-Tao Jiang, Yue Liao, Junwei Zhu, Jiangning Zhang, Ying Tai, Chengjie Wang, and Shuicheng Yan. Fera: Frequency-energy con- strained routing for effective diffusion adaptation fine- tuning.arXiv preprint arXiv:2511.17979, 2025. 3

    Pith/arXiv arXiv 2025

  72. [72]

    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. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821–14831, 2021. 2, 5, 4

    2021

  73. [73]

    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. InCVPR, 2022. 2, 5

    2022

  74. [74]

    Learning enriched features for fast image restoration and enhancement.TPAMI, 2022

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for fast image restoration and enhancement.TPAMI, 2022. 2, 5

    2022

  75. [75]

    Density-aware single image de-raining using a multi-stream dense network

    He Zhang and Vishal M Patel. Density-aware single image de-raining using a multi-stream dense network. InCVPR,

  76. [76]

    Ingredient-oriented multi- degradation learning for image restoration

    Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. Ingredient-oriented multi- degradation learning for image restoration. InCVPR, 2023. 5

    2023

  77. [77]

    R1-vl: Learn- ing to reason with multimodal large language models via step-wise group relative policy optimization.arXiv preprint arXiv:2503.12937, 2025

    Jingyi Zhang, Jiaxing Huang, Huanjin Yao, Shunyu Liu, Xikun Zhang, Shijian Lu, and Dacheng Tao. R1-vl: Learn- ing to reason with multimodal large language models via step-wise group relative policy optimization.arXiv preprint arXiv:2503.12937, 2025. 3

    Pith/arXiv arXiv 2025

  78. [78]

    Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising.TIP, 2017

    Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising.TIP, 2017. 2, 3, 5

    2017

  79. [79]

    Learning deep CNN denoiser prior for image restoration

    Kai Zhang, Wangmeng Zuo, Shuhang Gu, and Lei Zhang. Learning deep CNN denoiser prior for image restoration. In CVPR, 2017. 5

    2017

  80. [80]

    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. TIP, 2018. 5

    2018

Showing first 80 references.