Compression-Aware One-Step Diffusion Model for JPEG Artifact Removal

Jinpei Guo; Wenbo Li; Yong Guo; Yulun Zhang; Zheng Chen

REVIEW 3 major objections 6 minor 56 references

Compression-Aware One-Step Diffusion Model for JPEG Artifact Removal

T0 review · 3 major / 6 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 CODiff shows that JPEG artifact removal can succeed in one diffusion sampling step if the model is conditioned on compression-aware embeddings.

desk verdict Solid one-step diffusion for JPEG artifact removal, but the evaluation leans on a training-loss metric and omits PSNR/SSIM, so 'surpasses' is not yet proven. read the letter →

arxiv 2502.09873 v3 pith:7SFTOXFR submitted 2025-02-14 cs.CV

Jinpei Guo , Zheng Chen , Wenbo Li , Yong Guo , Yulun Zhang This is my paper · ORCID

classification cs.CV

keywords JPEGartifactremovalone-stepdiffusionmodelcompression-awarevisualembedderqualityfactorpredictionduallearninglatentimagerestorationperceptualloss

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

CODiff sets out to show that JPEG artifact removal can be done in a single diffusion sampling step without sacrificing perceptual quality, even at severe compression levels. The paper's central claim is that injecting compression-aware embeddings into a one-step latent diffusion model lets it outperform both multi-step diffusion restorers and previous one-step models on heavily compressed images. A supporting claim is that a dual training objective—explicit quality-factor prediction plus implicit reconstruction—gives the embedder enough compression understanding to generalize to quality factors never seen in training. If true, this matters because diffusion-based restoration becomes practical for real-time or high-throughput deployment while still handling the hardest JPEG degradations.

What carries the argument

The load-bearing object is the compression-aware visual embedder (CaVE), a UNet encoder whose feature vectors serve as JPEG compression priors for the denoising network. CaVE is trained with a dual strategy: a lightweight QF predictor supervises the averaged embedding with an L1 loss (explicit learning), while a UNet decoder reconstructs the high-quality image from the same embeddings with an L1 loss (implicit learning). The same embeddings replace the textual prompt in the latent diffusion generator, which is fine-tuned with LoRA and trained with L2, DISTS, and GAN losses. The argument is carried by this conditioning path: compression prior in, better single-step estimate out.

What would settle it

Test CODiff on JPEG images produced by encoders or pipelines outside its training distribution—for example, camera-ISP JPEGs or images compressed with QF below 8 or with chroma subsampling settings different from the synthetic pairs—and compare LPIPS and MUSIQ against OSEDiff* and a 50-step DiffBIR*. If the compression-prior advantage shrinks or reverses on those images, the generalization claim is falsified.

Watch

Extended reading notes

Core claim

On the paper's terms, CODiff is a compression-aware one-step diffusion model: given a JPEG-compressed image, its VAE encoder produces a latent, CaVE produces compression prior embeddings, and a single denoising step estimates the clean latent, which the decoder turns into the restored image. The discovered point is that the compression prior is the difference-maker. CaVE is trained so that its embeddings both predict the JPEG quality factor (explicit learning) and reconstruct the original image (implicit learning), and the paper reports that this dual objective separates compression levels in embedding space better than QF prediction alone, including for unseen QF 1 and 5. With that conditioning, CODiff reports lower LPIPS and DISTS and higher MUSIQ, MANIQA, and CLIPIQA than FBCNN, JDEC, PromptCIR, DiffBIR*, SUPIR, and OSEDiff* on LIVE-1, Urban100, and DIV2K-Val, using one sampling step, 1.00 G parameters, and 9.46 T MACs.

Load-bearing premise

The method's reported advantage rests on the assumption that JPEG images synthesized in training by compressing clean images with quality factors sampled from 8 to 95 represent the artifacts that appear in real test images, including at the unseen quality factors 1 and 5.

Editorial extensions

If this is right

A single sampling step is enough for competitive JPEG artifact removal, so diffusion-based restoration can operate at interactive speeds on 1024×1024 images.
Compression-level awareness generalizes beyond training quality factors: CaVE embeddings trained with dual learning separate unseen QF 1 and 5 clusters, improving restoration at extreme compression.
Compression priors can replace text prompts in latent diffusion restoration: CaVE outperforms empty, learnable, and DAPE prompt embeddings in the paper's ablations.
The combination of DISTS perceptual loss and GAN loss is what turns the one-step model from pixel-accurate but smooth to perceptually realistic outputs.
Because CODiff needs fewer parameters than OSEDiff, DiffBIR, and SUPIR, compression-aware conditioning gives a more efficient route than auxiliary prompt extractors or ControlNet.

Reading between the lines

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

The CaVE recipe should transfer to other block-based codecs such as WebP, AVIF, or HEIC by swapping the QF predictor for each codec's quality parameter, which is a direct testable extension of the same dual-learning idea.
If real JPEG artifacts differ from the synthetic training pairs, the likely failure point will be the scalar QF: real images often carry encoder-specific quantization tables, so a richer compression signature such as the quantization matrix may be needed, and the paper's current design bets that QF plus reconstruction is enough.
The reported MUSIQ and CLIPIQA gains at unseen QF 1 and 5 suggest the model is not merely memorizing quality levels but learning a continuous degradation axis; that axis could be probed directly by measuring artifact removal performance at finer QF steps.

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.

Desk Editor's Note

read the letter

Short version: this is a competent, incremental engineering paper that does something genuinely new—one-step diffusion for JPEG artifact removal—and supports it with fairly thorough ablations. But the headline claim of superiority is only partially supported: the full-reference metric they lean on (DISTS) is also in their training loss, and PSNR/SSIM are absent, which is a real problem for a restoration paper. I would send it to review, but the authors need to fix the evaluation and soften the 'consistently outperforms' language.

What is actually new: CODiff brings one-step latent diffusion to JPEG artifact removal, with a compression-aware visual embedder (CaVE) trained by a dual objective—explicit QF prediction plus implicit reconstruction. That combination is new relative to OSEDiff, FBCNN, and PromptCIR. The efficiency numbers are concrete: one step, 9.46 T MACs, 0.57 s at 1024×1024, versus 188.24 T MACs and 50.81 s for DiffBIR. Ablations support the CaVE design, and they retrain DiffBIR and OSEDiff on their own data for fair comparison, which is better than most papers in this area.

Soft spots, in order of importance. First, DISTS appears as a training loss (Eq. 7) with weight 1, and then as the principal full-reference evaluation metric. Some of the large DISTS gains are therefore measuring optimization success, not artifact-removal quality. LPIPS is not in the loss, but it is also a learned perceptual metric. Second, PSNR and SSIM are missing from every table. For a task called 'artifact removal,' those are the standard measures of whether content is actually recovered rather than regenerated. Without them, a GAN-trained model can score well on perceptual and no-reference metrics while inventing texture. Third, the paper says CODiff 'consistently outperforms,' but Table 1 shows PromptCIR beating CODiff on MUSIQ at QF=20 on both Urban100 and DIV2K-val. That is an overstatement. Fourth, smaller issues: the diffusion timestep TL in Eq. (1) is never specified; Sec. 3.2.2 confuses the symbols zL and cL; and the figure 7 caption area contains garbled Unicode garbage that should be cleaned up.

The weakest assumption, as the stress-test notes, is the synthetic JPEG training with QF labels. The generalization to unseen QF 1 and 5 is a genuine test, and they pass it, but real-world JPEG artifacts could differ. That said, this is a reasonable engineering contribution with addressable flaws, not a broken paper. Who is it for: people working on fast diffusion restoration or compression artifact removal. I would engage with it in review.

Referee Report

3 major / 6 minor

Summary. The paper proposes CODiff, a one-step latent diffusion model for JPEG artifact removal. It introduces a compression-aware visual embedder (CaVE) trained with a dual objective: explicit quality-factor prediction and implicit high-quality image reconstruction. In a second stage, the CaVE embeddings condition a Stable Diffusion UNet that is fine-tuned with LoRA using L2, DISTS, and GAN losses. The method is evaluated on LIVE-1, Urban100, and DIV2K-Val against FBCNN, JDEC, PromptCIR, DiffBIR*, SUPIR, and OSEDiff*, reporting LPIPS, DISTS, MUSIQ, MANIQA, CLIPIQA, and complexity. The authors claim state-of-the-art performance with a single sampling step and reduced computational cost relative to multi-step diffusion baselines.

Significance. If the reported results are substantiated, CODiff would be a practically valuable contribution: it demonstrates that compression-aware conditioning can be combined with one-step diffusion, achieving low perceptual error at 9.46 T MACs and releasing code. The dual-learning idea for CaVE is interesting, and the evaluation on unseen quality factors (QF=1 and 5) is a genuine generalization test. However, the evidence for the central superiority claim is incomplete because standard fidelity metrics are omitted and one evaluation metric is also a training loss, and some of the reported comparisons do not support the unqualified claim of consistent improvement.

major comments (3)

[§4.1 / Eq. (7); Tables 1–5] The headline quantitative evidence is weakened by a train/test metric overlap. DISTS appears both as a training loss (Eq. (7), with λ_D = 1, Sec. 4.1) and as one of the two full-reference evaluation metrics in Tables 1–5, while PSNR and SSIM—the standard fidelity metrics for JPEG artifact removal—are not reported anywhere. Since the generator is also trained with a GAN loss (Eqs. (8)–(9)), the large DISTS/LPIPS gains (e.g., LIVE-1 QF=5: DISTS 0.1121 vs 0.1653 for OSEDiff*) could partly reflect direct optimization of the evaluation metric and synthesis of plausible texture rather than faithful artifact removal. Without PSNR/SSIM or another fidelity measure, the abstract's claim that CODiff 'surpasses recent leading methods' for JPEG artifact removal is not fully supported.
[§4.2, Table 1(b,c)] The text states that CODiff 'consistently outperforms competing methods' across a diverse range of metrics, but Table 1 shows exceptions at QF=20. On Urban100, CODiff's MUSIQ is 72.63, below PromptCIR's 73.47, and its LPIPS 0.0753 is above PromptCIR's 0.0739. On DIV2K-val, CODiff's MUSIQ 66.66 is below JDEC's 69.90 and DiffBIR's 67.06. The abstract's unqualified claim of superiority is therefore too strong; the claims should be restricted to the metrics and quality factors where the advantage actually holds.
[§4.2 / Fig. 5] The paper emphasizes generalization to unseen, highly compressed images, and Fig. 5 includes a QF=1 example, but Table 1 reports quantitative results only for QF=5, 10, and 20. Since QF=1 lies outside the training range (8–95) and is the extreme case highlighted in the motivation, the absence of quantitative metrics at QF=1 leaves the generalization claim supported only by t-SNE plots and QF-prediction MSE (Figs. 4 and 7) rather than by restoration quality. Please add numerical results at QF=1 or explicitly scope the claim.

minor comments (6)

[Fig. 7] The caption contains raw '/uni00000014/...' tokens, apparently a font or PDF encoding artifact; the caption should be regenerated.
[Eq. (6) / Sec. 4.1] The symbol λ is used for the reconstruction weight (λ = 1,000) while λ_D and λ_G denote the DISTS and GAN loss weights; rename one of these to avoid ambiguity.
[Sec. 3.2.2] The QF-prediction formula first writes P(¯c_L; ϕ) and then defines ¯z_L as the averaged embedding; the notation should be unified to ¯z_L.
[Table 1] The table uses 's=50' and 's=1' without defining s in the caption; define it as the number of sampling steps.
[Table 5] The checkmark columns are visually ambiguous: the third row appears to denote LDISTS + LG but the alignment is unclear. Specify which losses are active in each row.
[References] Reference [33] lists only 'H Sheikh'; provide full author names.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CODiff's headline comparisons are held-out benchmark evaluations, including unseen QF=1 and 5; the DISTS loss/metric overlap is an evaluation caveat, not a derivation loop.

full rationale

The paper's derivation chain is self-contained against external benchmarks. CaVE is trained with a QF-regression loss (Eq. 4) and an L1 reconstruction loss (Eq. 5), and the final one-step diffusion generator is trained with L2 + DISTS + GAN losses (Eqs. 7-10). The reported results on LIVE-1, Urban100, and DIV2K-val are measured on held-out images, and QF=1 and QF=5 are explicitly outside the training range QF∈[8,95] (Sec. 4.1), so the central claim does not reduce to a fitted parameter or to a self-citation. The only notable overlap is that DISTS appears both as a training loss (Eq. 7) and as a reported metric (Table 1), which weakens DISTS as independent evidence but is not a circularity of the derivation; LPIPS, MUSIQ, MANIQA, and CLIPIQA are not optimized. Self-citations such as Refs. [11] and [23] are background context and are not load-bearing for the central claim. No step was found in which a predicted quantity is defined in terms of the target result.

Assumptions & free parameters 6 free parameters · 4 assumptions · 1 invented entities

The central claim rests on a standard diffusion pipeline plus several hand-set losses and a synthetic degradation model. The biggest external input is the pretrained Stable Diffusion prior; the biggest internal choices are the loss weights and the QF range. No hidden physical constants are introduced.

free parameters (6)

lambda (implicit loss weight) = 1000
Chosen by hand (Sec. 4.1) to balance QF prediction and reconstruction in CaVE; no sensitivity analysis is provided.
lambda_D (DISTS weight) = 1
Chosen by hand in Sec. 4.1; ablation in Tab. 5 shows DISTS is the main driver of perceptual gains.
lambda_G (GAN loss weight) = 0.005
Chosen by hand in Sec. 4.1; ablation shows a small additional gain.
LoRA rank = 16
Set in Sec. 4.1 for fine-tuning the Stable Diffusion UNet.
Diffusion timestep T_L = not specified
Eq. (1) uses a 'predefined diffusion timestep' but the value is never given; reproducibility depends on it.
Training QF range = 8 to 95
Sampling range for synthetic compression in Sec. 4.1; unseen QFs 1 and 5 are used to claim generalization.

assumptions (4)

standard math Latent diffusion forward and reverse equations (z_t = sqrt(alpha_bar_t) z + sqrt(1-alpha_bar_t) epsilon, etc.)
Used as the foundation of the one-step OSD formulation in Sec. 3.1.
domain assumption Pretrained Stable Diffusion VAE and UNet provide useful generative priors for image restoration
The whole method builds on adapting SD via LoRA; no evidence is given that weaker priors would work.
domain assumption Synthetic JPEG compression with QF in [8, 95] represents the JPEG artifact distribution at test time
Training and evaluation are both on synthetic JPEG; no real-world compressed images are used.
domain assumption Perceptual metrics (LPIPS, DISTS) and no-reference IQA (MUSIQ, MANIQA, CLIPIQA) capture restoration quality better than PSNR/SSIM
Only these metrics are reported; PSNR and SSIM are omitted.

invented entities (1)

CaVE (compression-aware visual embedder) independent evidence
purpose: Extract JPEG compression priors from the input image and inject them as prompt embeddings into the one-step diffusion UNet (Sec. 3.2).
This is a new network module introduced by the paper; its contribution is supported by ablations (Tab. 3, 4) and the t-SNE and QF prediction visualizations, so it is not an unfalsifiable postulate.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Compression-Aware One-Step Diffusion Model for JPEG Artifact Removal." pith.science (2026). https://pith.science/paper/7SFTOXFR

@misc{pith2026250209873,
  author       = {Pith},
  title        = {Pith review of: Compression-Aware One-Step Diffusion Model for JPEG Artifact Removal},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7SFTOXFR}},
  note         = {Machine review of arXiv:2502.09873}
}

read the original abstract

Diffusion models have demonstrated remarkable success in image restoration tasks. However, their multi-step denoising process introduces significant computational overhead, limiting their practical deployment. Furthermore, existing methods struggle to effectively remove severe JPEG artifact, especially in highly compressed images. To address these challenges, we propose CODiff, a compression-aware one-step diffusion model for JPEG artifact removal. The core of CODiff is the compression-aware visual embedder (CaVE), which extracts and leverages JPEG compression priors to guide the diffusion model. We propose a dual learning strategy that combines explicit and implicit learning. Specifically, explicit learning enforces a quality prediction objective to differentiate low-quality images with different compression levels. Implicit learning employs a reconstruction objective that enhances the model's generalization. This dual learning allows for a deeper and more comprehensive understanding of JPEG compression. Experimental results demonstrate that CODiff surpasses recent leading methods in both quantitative and visual quality metrics. The code is released at https://github.com/jp-guo/CODiff.

Figures

Figures reproduced from arXiv: 2502.09873 by the authors.

**Figure 2.** An illustration of the explicit learning approach with our [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Overview of our proposed CODiff. In the first stage, we train our compression-aware visual embedder (CaVE) via a dual [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Visualization of JPEG prior embeddings from CaVE under different training objectives. In (a) and (c), CaVE is trained using only [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: Visual comparison on different quality factors (QF=1, 5 and 10). DiffBIR* and OSEDiff* are retrained as references. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 7.** Figure 7: Comparison of CaVE’s QF prediction mean squared [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 39 canonical work pages

[1]

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. In CVPRW, 2017. 5

work page 2017
[2]

Diffusion models beat gans on image synthesis

Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 1

work page 2021
[3]

Image quality assessment: Unifying structure and texture similarity

Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE TPAMI, 2020. 5

work page 2020
[4]

Learning a deep convolutional network for image super-resolution

Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang. Learning a deep convolutional network for image super-resolution. In ECCV, 2014. 2

work page 2014
[5]

Compression artifacts reduction by a deep convolu- tional network

Chao Dong, Yubin Deng, Chen Change Loy, and Xiaoou Tang. Compression artifacts reduction by a deep convolu- tional network. In ICCV, 2015. 1, 2

work page 2015
[6]

Quantization guided jpeg artifact correction

Max Ehrlich, Larry Davis, Ser-Nam Lim, and Abhinav Shri- vastava. Quantization guided jpeg artifact correction. In ECCV, 2020. 1, 2

work page 2020
[7]

Deep generative adversarial compression artifact removal

Leonardo Galteri, Lorenzo Seidenari, Marco Bertini, and Al- berto Del Bimbo. Deep generative adversarial compression artifact removal. In ICCV, 2017. 2

work page 2017
[8]

Deep universal generative adversarial compression artifact removal

Leonardo Galteri, Lorenzo Seidenari, Marco Bertini, and Alberto Del Bimbo. Deep universal generative adversarial compression artifact removal. IEEE TMM, 2019. 2

work page 2019

Show all 56 references

[9]

Generative adversarial networks

Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Commu- nications of the ACM, 2020. 5

2020
[10]

Building dual-domain rep- resentations for compression artifacts reduction

Jun Guo and Hongyang Chao. Building dual-domain rep- resentations for compression artifacts reduction. In ECCV,
[11]

Oscar: One- step diffusion codec across multiple bit-rates

Jinpei Guo, Yifei Ji, Zheng Chen, Kai Liu, Min Liu, Wang Rao, Wenbo Li, Yong Guo, and Yulun Zhang. Oscar: One- step diffusion codec across multiple bit-rates. arXiv preprint arXiv:2505.16091, 2025. 3

2025
[12]

Jdec: Jpeg decoding via enhanced continuous cosine coefficients

Woo Kyoung Han, Sunghoon Im, Jaedeok Kim, and Ky- ong Hwan Jin. Jdec: Jpeg decoding via enhanced continuous cosine coefficients. In CVPR, 2024. 1, 2, 6, 7

2024
[13]

Denoising diffu- sion probabilistic models

Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models. NeurIPS, 2020. 1

2020
[14]

Lora: Low-rank adaptation of large language models

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3, 5, 6

2021 arXiv
[15]

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. 2, 5

2015
[16]

Towards flexible blind jpeg artifacts removal

Jiaxi Jiang, Kai Zhang, and Radu Timofte. Towards flexible blind jpeg artifacts removal. In ICCV, 2021. 1, 2, 4, 6, 7

2021
[17]

Autodir: Automatic all-in-one image restoration with latent diffusion

Yitong Jiang, Zhaoyang Zhang, Tianfan Xue, and Jinwei Gu. Autodir: Automatic all-in-one image restoration with latent diffusion. In ECCV, 2024. 3

2024
[18]

Musiq: Multi-scale image quality transformer

Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In ICCV, 2021. 5

2021
[19]

Agar- net: Adaptively gated jpeg compression artifacts removal network for a wide range quality factor

Yoonsik Kim, Jae Woong Soh, and Nam Ik Cho. Agar- net: Adaptively gated jpeg compression artifacts removal network for a wide range quality factor. IEEE Access, 2020. 2

2020
[20]

Adam: A method for stochastic opti- mization

Diederik P Kingma. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980, 2014. 5

2014 arXiv
[21]

Promptcir: Blind compressed image restoration with prompt learning

Bingchen Li, Xin Li, Yiting Lu, Ruoyu Feng, Mengxi Guo, Shijie Zhao, Li Zhang, and Zhibo Chen. Promptcir: Blind compressed image restoration with prompt learning. arXiv preprint arXiv:2404.17433, 2024. 1, 2, 5, 6, 7

2024 arXiv
[22]

Learning a single model with a wide range of quality factors for jpeg image artifacts removal

Jianwei Li, Yongtao Wang, Haihua Xie, and Kai-Kuang Ma. Learning a single model with a wide range of quality factors for jpeg image artifacts removal. IEEE TIP, 2020. 2

2020
[23]

Distillation-free one-step diffusion for real-world image super-resolution

Jianze Li, Jiezhang Cao, Zichen Zou, Xiongfei Su, Xin Yuan, Yulun Zhang, Yong Guo, and Xiaokang Yang. Distillation-free one-step diffusion for real-world image super-resolution. arXiv preprint arXiv:2410.04224, 2024. 1

2024 arXiv
[24]

Lsdir: A large scale dataset for image restoration

Yawei Li, Kai Zhang, Jingyun Liang, Jiezhang Cao, Ce Liu, Rui Gong, Yulun Zhang, Hao Tang, Yun Liu, Denis Deman- dolx, et al. Lsdir: A large scale dataset for image restoration. In CVPR, 2023. 5

2023
[25]

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. In ICCV, 2021. 1, 2

2021
[26]

Diff- bir: Toward blind image restoration with generative diffusion prior

Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diff- bir: Toward blind image restoration with generative diffusion prior. In ECCV, 2024. 1, 3, 6, 7, 8

2024
[27]

Decoupled weight decay regularization

I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5

2017 arXiv
[28]

Pytorch: An imperative style, high-performance deep learning library

Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 5

2019
[29]

Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis

Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 3, 6

2023 arXiv
[30]

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. In CVPR, 2022. 1, 3, 6, 7

2022
[31]

U-net: Convolutional networks for biomedical image segmentation

Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4, 5

2015
[32]

Photorealistic text-to-image diffusion models with deep language understanding

Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 1, 3 9

2022
[33]

Live image quality assessment database release 2

H Sheikh. Live image quality assessment database release 2. http://live. ece. utexas. edu/research/quality, 2005. 5

2005
[34]

Denoising diffusion implicit models

Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1

2010 arXiv
[35]

Score-based generative modeling through stochastic differential equa- tions

Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equa- tions. arXiv preprint arXiv:2011.13456, 2020. 1

2011 arXiv
[36]

Ntire 2017 challenge on single image super-resolution: Methods and results

Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming- Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In CVPRW,

2017
[37]

Visualizing data using t-sne

Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 2008. 2, 4

2008
[38]

The jpeg still picture compression stan- dard

Gregory K Wallace. The jpeg still picture compression stan- dard. Communications of the ACM, 1991. 1, 2, 6

1991
[39]

Ex- ploring clip for assessing the look and feel of images

Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Ex- ploring clip for assessing the look and feel of images. In AAAI, 2023. 5

2023
[40]

Exploiting diffusion prior for real-world image super-resolution

Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. IJCV, 2024. 3

2024
[41]

Jpeg artifacts removal via compression quality ranker-guided networks

Menglu Wang, Xueyang Fu, Zepei Sun, and Zheng-Jun Zha. Jpeg artifacts removal via compression quality ranker-guided networks. In IJCAI, 2021. 2

2021
[42]

Jpeg artifacts removal via contrastive representation learning

Xi Wang, Xueyang Fu, Yurui Zhu, and Zheng-Jun Zha. Jpeg artifacts removal via contrastive representation learning. In ECCV, 2022. 4

2022
[43]

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

Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image super- resolution in a single step. In CVPR, 2024. 3, 5

2024
[44]

One-step effective diffusion network for real-world image super-resolution

Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. arXiv preprint arXiv:2406.08177, 2024. 1, 3, 5, 6, 7, 8

2024 arXiv
[45]

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

Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In CVPR, 2024. 1, 3, 7, 8

2024
[46]

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

Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniqa: Multi-dimension attention network for no-reference image quality assessment. In CVPR, 2022. 5

2022
[47]

Pixel-aware stable diffusion for realistic im- age super-resolution and personalized stylization

Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic im- age super-resolution and personalized stylization. In ECCV,
[48]

Im- proved distribution matching distillation for fast image syn- thesis

Tianwei Yin, Micha ¨el Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Im- proved distribution matching distillation for fast image syn- thesis. arXiv preprint arXiv:2405.14867, 2024. 3, 5

2024 arXiv
[49]

One-step diffusion with distribution matching distillation

Tianwei Yin, Micha ¨el Gharbi, Richard Zhang, Eli Shecht- man, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In CVPR, 2024. 3, 5

2024
[50]

Scaling up to excellence: Practicing model scaling for photo- realistic image restoration in the wild

Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photo- realistic image restoration in the wild. In CVPR, 2024. 1, 3, 6, 7, 8

2024
[51]

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

Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Resshift: Efficient diffusion model for image super- resolution by residual shifting. NeurIPS, 2024. 3

2024
[52]

Adding conditional control to text-to-image diffusion models

Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 7

2023
[53]

The unreasonable effectiveness of deep features as a perceptual metric

Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5

2018
[54]

Dmcnn: Dual-domain multi-scale convolutional neural net- work for compression artifacts removal

Xiaoshuai Zhang, Wenhan Yang, Yueyu Hu, and Jiaying Liu. Dmcnn: Dual-domain multi-scale convolutional neural net- work for compression artifacts removal. In ICIP, 2018. 2

2018
[55]

Residual non-local attention networks for image restora- tion

Yulun Zhang, Kunpeng Li, Kai Li, Bineng Zhong, and Yun Fu. Residual non-local attention networks for image restora- tion. arXiv preprint arXiv:1903.10082, 2019. 1, 2

1903 arXiv
[56]

Implicit dual-domain convolutional network for robust color image compression artifact reduction

Bolun Zheng, Yaowu Chen, Xiang Tian, Fan Zhou, and Xuesong Liu. Implicit dual-domain convolutional network for robust color image compression artifact reduction. IEEE TCSVT, 2019. 2 10

2019

Pith tools

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

[1] [1]

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. In CVPRW, 2017. 5

work page 2017

[2] [2]

Diffusion models beat gans on image synthesis

Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 1

work page 2021

[3] [3]

Image quality assessment: Unifying structure and texture similarity

Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE TPAMI, 2020. 5

work page 2020

[4] [4]

Learning a deep convolutional network for image super-resolution

Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang. Learning a deep convolutional network for image super-resolution. In ECCV, 2014. 2

work page 2014

[5] [5]

Compression artifacts reduction by a deep convolu- tional network

Chao Dong, Yubin Deng, Chen Change Loy, and Xiaoou Tang. Compression artifacts reduction by a deep convolu- tional network. In ICCV, 2015. 1, 2

work page 2015

[6] [6]

Quantization guided jpeg artifact correction

Max Ehrlich, Larry Davis, Ser-Nam Lim, and Abhinav Shri- vastava. Quantization guided jpeg artifact correction. In ECCV, 2020. 1, 2

work page 2020

[7] [7]

Deep generative adversarial compression artifact removal

Leonardo Galteri, Lorenzo Seidenari, Marco Bertini, and Al- berto Del Bimbo. Deep generative adversarial compression artifact removal. In ICCV, 2017. 2

work page 2017

[8] [8]

Deep universal generative adversarial compression artifact removal

Leonardo Galteri, Lorenzo Seidenari, Marco Bertini, and Alberto Del Bimbo. Deep universal generative adversarial compression artifact removal. IEEE TMM, 2019. 2

work page 2019