Pith. sign in

REVIEW 4 major objections 4 minor 39 references

Restoring Real-World Images with an Internal Detail Enhancement Diffusion Model

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

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

Pith's one-line read Fine-tuning a frozen diffusion model on its own denoising estimates preserves detail in restored real-world images and enables object-level color control from text prompts.

desk verdict A plausible but under-specified method: the IIDE self-regularization is asserted rather than derived, and the actual training loss does not implement the claimed consistency constraint. read the letter →

arxiv 2505.18674 v2 pith:HB7DYKYW submitted 2025-05-24 cs.CV cs.AI

Peng Xiao , Hongbo Zhao , Yijun Wang , Jianxin Lin This is my paper · ORCID
classification cs.CVcs.AI
keywords real-worldimagerestorationtext-guidedold-photocolorizationsuper-resolutiondiffusionmodelsdetailpreservationControlNetDDIM
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that a frozen text-to-image diffusion model can be turned into a high-fidelity restorer of real-world degraded images by fine-tuning only a small ControlNet branch with a self-regularization scheme called Internal Image Detail Enhancement (IIDE). IIDE replaces the manually designed degradation operators used by prior conditional diffusion methods with the model's own estimate of a clean image, computed mid-denoising by DDIM, and trains the network to produce consistent results whether it is conditioned on the original degraded photo or on that estimate. The authors report that this makes restored old photos and super-resolved images perceptually superior to recent state-of-the-art methods, with higher CLIPIQA and MUSIQ scores and lower FID, at a modest cost in PSNR and SSIM. The method retains text-guided editing ability, so a user can recolor specific objects in a restored photo. If true, the approach offers a way to restore old and low-quality images with detail fidelity and user control using an off-the-shelf generative prior rather than a task-specific model trained from scratch.

What carries the argument

The load-bearing object is the DDIM clean-image prediction $\tilde{x}_0 = D(\tilde{z}_0)$ computed from the noisy latent $z_{t-1}$ via Eq. 4, used as an automatic, step-dependent degradation proxy. IIDE treats this estimate as a stand-in for a realistic corrupted version of the true image and enforces, through mix-up training, that $p_\theta(z_{t-1} | z_t, C, I_{lq}) \approx p_\theta(z_{t-1} | z_t, C, \tilde{x}_0)$; this self-regularization is what preserves detail. The trainable part is the ControlNet branch feeding the degraded image and, for old photos, a scratch mask into a frozen Stable Diffusion UNet, while the controllable feature transformation module from StableSR is added after training for the super-resolution fidelity-quality tradeoff.

What would settle it

Take synthetic pairs of clean and degraded images with known scratch masks, compute the DDIM estimate $\tilde{x}_0$ from intermediate latents, and measure the perceptual distance (e.g., LPIPS) between $\tilde{x}_0$ and the actual degraded image; if the distance is large precisely in the scratched regions where the mask lies, the equivalence in Eq. 9 fails and IIDE would be conditioning the network on content unrelated to the true degradation.

Watch

Extended reading notes

Core claim

The central claim is that detail loss in diffusion-based restoration, which stems from the stochasticity of the generative process, can be suppressed by an internal consistency constraint: during fine-tuning, the model is asked to produce the same denoising trajectory whether its visual condition is the corrupted input $I_{lq}$ or the DDIM-predicted clean estimate $\tilde{x}_0$ derived from the current intermediate latent using Eq. 4. Because $\tilde{x}_0$ is generated automatically at every step, the method avoids hand-designing degradation operators, and the condition that the two trajectories coincide acts as a self-regularization that forces the latent $z_{t-1}$ to retain the details of $I_{lq}$. Trained this way, with a mix-up probability $p_{\text{iide}} = 0.5$ on top of Stable Diffusion 2.1 and ControlNet, the model outperforms BrOldPho, DeOldify, StableSR, SinSR, DiffIR, and other methods on perceptual metrics for old photo restoration and super-resolution, and it supports text-guided object-level colorization.

Load-bearing premise

The load-bearing premise is that the denoiser's automatically estimated clean image from a mid-denoising latent is a faithful degraded version of the true photo, so that conditioning on it is interchangeable with conditioning on the real corrupted input; if that estimate is not faithful, the detail-preserving constraint can amplify errors instead of details.

Editorial extensions

If this is right

  • On old photo restoration, the IIDE-trained model improves CLIPIQA by 53.6% over BrOldPho and MUSIQ by 29.5% over BrOldPho plus DeOldify, while lowering FID by 18.4%.
  • On image super-resolution, the method exceeds StableSR in CLIPIQA and MUSIQ on DIV2K (+3.6% and +1.6%), RealSR (+7.4% and +0.3%), and DRealSR (+5.5% and +7.3%).
  • Text-guided restoration works at object level: changing the prompt changes object colors while keeping the content, as demonstrated on both gray synthetic photos and real old photos.
  • The method fine-tunes only the ControlNet branch of a frozen pre-trained diffusion model, so it avoids training a diffusion model from scratch and keeps the number of trainable parameters low.

Reading between the lines

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

  • The method's robustness likely depends on the training data covering the same kinds of degradation that the automatic clean-image estimate produces; on heavily localized or unseen corruptions, the self-regularization might reinforce artifacts instead of true detail recovery.
  • Because the reported gains sit in perceptual metrics while PSNR and SSIM drop slightly, the approach suits human-viewing applications; for medical or satellite imagery, an explicit fidelity term would likely be needed to prevent hallucinated textures.
  • The same internal-consistency recipe could extend to video restoration or deblurring by producing the automatic clean estimate from temporally coherent denoising paths, though the paper does not test these settings.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The manuscript proposes Internal Image Detail Enhancement (IIDE), a fine-tuning strategy that adapts a frozen Stable Diffusion model, via a trainable ControlNet, to real-world image restoration (old-photo restoration, super-resolution, text-guided colorization). The key idea is to use the DDIM estimate x~0 of the clean image from an intermediate noisy latent as a surrogate degraded condition, and to randomly replace the original low-quality conditioning image Ilq with x~0 during training. Experiments on DIV2K, RealSR, and DRealSR report perceptual-metric improvements (CLIPIQA, MUSIQ) over several baselines, with acknowledged PSNR/SSIM trade-offs, plus qualitative demonstrations of text-guided object-level color editing.

Significance. If validated, the contribution is practically relevant: it achieves perceptual improvements on multiple restoration tasks using a single frozen generative prior, trains only a ControlNet, and enables text-guided color editing on real old photos. The ablation 'w/ IIDE' vs 'w/o IIDE' in Tables I and IV suggests IIDE has a non-trivial effect. However, the theoretical derivation of IIDE is currently informal, the implemented training loss does not match the claimed consistency constraint, and the quantitative evidence lacks error bars or statistical tests. These gaps prevent the paper from establishing a reliable new mechanism, though the empirical direction is promising.

major comments (4)
  1. [Section III-B, Eq. (9)] The derivation treats the DDIM estimate x~0 = D(z~0) from Eq. (4) as 'a degradation version of Ihq,' but Eq. (4) is a deterministic denoising reconstruction, not a degradation operator. The statement q(zt|Ilq) ≈ q(zt|x) for large t is about marginal distributions of noisy latents; it does not imply the specific x~0 computed from latents conditioned on Ilq is a valid degradation of Ihq. At training time x~0 is obtained from latents initialized from ground-truth Ihq, whereas at inference it must be computed from latents driven by Ilq and noise; no argument is given that these distributions match. Since this substitution is the foundation of the IIDE mechanism and the claim that IIDE 'ensures' detail-preserved images, it is load-bearing and needs either a rigorous justification or a reformulation of the claim.
  2. [Section III-C] The training loss described in Section III-C is simply the standard diffusion loss of Eq. (2) with random condition replacement: with probability piide the condition is Ilq, otherwise it is x~0. There is no pairwise consistency term that enforces pθ(zt−1|zt,C,Ilq) ≈ pθ(zt−1|zt,C,x~0), so the self-regularization promised by Eq. (9) is not actually optimized. As implemented, IIDE is condition augmentation rather than a consistency constraint; the authors should either add an explicit consistency loss or revise the claims accordingly.
  3. [Tables I, II, IV] All quantitative claims of superiority are based on single-run aggregated metrics without error bars, confidence intervals, or statistical significance tests. This is particularly important because the method accepts lower PSNR/SSIM (e.g., Table II: PSNR 31.84 vs 33.46 for DiffIR; Table I: PSNR 29.11 vs 30.20 for BrOldPho+DeOldify) and the reported perceptual gains (e.g., CLIPIQA 0.7046 vs 0.6804 for StableSR in Table II) may be within run-to-run variation for stochastic diffusion sampling. The claim 'significantly outperforms state-of-the-art' in the Abstract and Section IV requires uncertainty quantification or a human study.
  4. [Section III-B, Eqs. (7)-(8)] The probability notation in this derivation is ill-defined: Eq. (7) integrates over x~0 as an image, while Eq. (8) expresses pθ(x~0|Ilq) through an integral over zt,zt−1 and uses p(˜x0|zt−1), even though Eq. (4) defines z~0 deterministically rather than as a distribution. Without a well-defined generative process for x~0, the step from Eq. (8) to Eq. (9) is not an inference but an assertion. This should be formalized or the derivation should be removed and replaced with a direct description of the algorithm.
minor comments (4)
  1. [Throughout] Several typos should be corrected, including 'an novel' and 'detial-preserving' in Section I, 'exihibited' in Section III, and 'Iterative Image Detail Enhancement' in Section IV-D2.
  2. [References] References [13] and [26] appear to be the same paper ('Exploiting diffusion prior for real-world image super-resolution'); one duplicate should be removed.
  3. [Section IV-D and Appendix] The colorization quantitative results appear only in Table IV in the appendix, while the main text claims object-level color control; consider moving a compact version into the main text for the reader.
  4. [Appendix D and E] The appendix section titles 'D. Appendix C: Additional Visual Comparisons for Old Photo Restoration' and 'E. Appendix D: Additional Visual Comparisons of Image Super-Resolution' are mislabeled and should be corrected to 'Appendix C' and 'Appendix D'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: IIDE is a training-time self-regularization heuristic, and the paper's reported gains are validated by external benchmark comparisons rather than by construction of the loss.

full rationale

The paper's central empirical claims--perceptual improvements in old-photo restoration, image super-resolution, and text-guided colorization--are evaluated against external baselines and reference-free metrics in Tables I-IV. IIDE is implemented as a condition-mixing training strategy: Section III-C replaces the vision condition Ilq with the model's own DDIM estimate x~0 with probability piide=0.5 and optimizes the standard diffusion loss of Eq. 2. The claimed detail-preservation 'guarantee' is a heuristic interpretation of the self-regularization, not a derived theorem, and its effect is shown by ablation (Ours w/ IIDE vs Ours w/o IIDE) on external metrics. The mathematically unsupported step is the identification of the DDIM clean estimate x~0 with a degradation of Ihq in Eq. 9 and the assertion that q(zt|Ilq) approximately equals q(zt|x) for large t; this is an unproven modeling assumption that may threaten correctness, but it does not make the reported predictions equivalent to the method's inputs by construction. There are no load-bearing self-citations, no fitted parameters renamed as predictions, and no imported uniqueness theorems. Hence no significant circularity is present.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical or architectural entities are introduced; IIDE is a training objective. The method adds a ControlNet branch and a scratch-mask condition, both standard components. The central claim rests on the empirical effectiveness of a self-regularization that is justified heuristically; the free parameters are minimal but loosely specified.

free parameters (3)
  • piide (mix-up probability) = 0.5
    Chosen by hand to balance training with the original image condition and the IIDE-replaced condition; affects training stability but no ablation is shown.
  • fidelity-quality tradeoff weight for feature transformation module = not reported
    In SR experiments, a controllable feature transformation module (after StableSR) is used to trade off fidelity and quality; the weight is not specified, so results may depend on an unreported setting.
  • old-photo degradation synthesis pipeline = not specified
    Training data for old photo restoration combines structured and unstructured degradations plus scratch masks, but the exact synthesis recipe is not described, leaving an uncontrolled variable.
assumptions (4)
  • domain assumption Pre-trained Stable Diffusion autoencoder and UNet provide a strong natural-image prior for restoration.
    The method freezes all Stable Diffusion weights and only trains ControlNet, relying on the prior to generate plausible details (Section III-C).
  • domain assumption The degradation operator phi_omega is content-preserving: restored output should be invariant to which degradation produced the input.
    Eq. 5 asserts Ire = G(Ilq, C) = G(phi_omega(Ihq), C), the basis for the consistency constraint.
  • ad hoc to paper The DDIM clean estimate from an intermediate noisy latent is a valid proxy for a real degradation of the ground truth.
    Section III-B replaces phi_omega(Ihq) with x~0 = D(z~0) in Eq. 9 without proof; the entire IIDE regularization depends on this equivalence.
  • domain assumption For large t, the forward noised latent of Ilq approximates that of Ihq.
    Used to justify q(zt|Ilq) ~ q(zt|x) in Section III-B, borrowed from standard diffusion theory.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Restoring Real-World Images with an Internal Detail Enhancement Diffusion Model." pith.science (2026). https://pith.science/paper/HB7DYKYW

@misc{pith2026250518674,
  author       = {Pith},
  title        = {Pith review of: Restoring Real-World Images with an Internal Detail Enhancement Diffusion Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HB7DYKYW}},
  note         = {Machine review of arXiv:2505.18674}
}
read the original abstract

Restoring real-world degraded images, such as old photographs or low-resolution images, presents a significant challenge due to the complex, mixed degradations they exhibit, such as scratches, color fading, and noise. Recent data-driven approaches have struggled with two main challenges: achieving high-fidelity restoration and providing object-level control over colorization. While diffusion models have shown promise in generating high-quality images with specific controls, they often fail to fully preserve image details during restoration. In this work, we propose an internal detail-preserving diffusion model for high-fidelity restoration of real-world degraded images. Our method utilizes a pre-trained Stable Diffusion model as a generative prior, eliminating the need to train a model from scratch. Central to our approach is the Internal Image Detail Enhancement (IIDE) technique, which directs the diffusion model to preserve essential structural and textural information while mitigating degradation effects. The process starts by mapping the input image into a latent space, where we inject the diffusion denoising process with degradation operations that simulate the effects of various degradation factors. Extensive experiments demonstrate that our method significantly outperforms state-of-the-art models in both qualitative assessments and perceptual quantitative evaluations. Additionally, our approach supports text-guided restoration, enabling object-level colorization control that mimics the expertise of professional photo editing.

Figures

Figures reproduced from arXiv: 2505.18674 by the authors.

Figure 1
Figure 1. Image Restoration under Various Real-World Degradations Given a degraded image with low quality, our method produces high-fidelity restorations and enables object-level color control when provided with text prompts. output space. While such approaches have achieved success on tasks such as image super-resolution [8] and image deblurring [10], the design of these approaches often focuses on one certain image degradat… view at source ↗
Figure 2
Figure 2. The framework of our proposed method, which is obtained by finetuning a pre-trained diffusion model with the Internal Image Detail Enhancement [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visual comparison of old photo restoration methods. Zoom-in for [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Visual comparisons of SR methods. Zoom-in for better details. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Visual comparison of text-guided image colorization methods. Zoom-in for better details. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Visual comparison of real-world old photo restoration. Zoom-in for better details. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Visual comparison of text-guided real-world old photo restoration. Zoom-in for better details. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Visual comparison of image super-resolution methods on synthetic dataset. Zoom-in for better details. [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Visual comparison of image super-resolution methods on real-world datasets. Zoom-in for better details. [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 36 canonical work pages

  1. [1]

    Bringing old photos back to life,

    Ziyu Wan, Bo Zhang, Dongdong Chen, Pan Zhang, Dong Chen, Jing Liao, and Fang Wen, “Bringing old photos back to life,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, 2020, pp. 2747–2757

    work page 2020

  2. [2]

    Pik-fix: Restoring and colorizing old photos,

    R. Xu, Z. Tu, Y . Du, et al., “Pik-fix: Restoring and colorizing old photos,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , 2023, pp. 1724–1734

    work page 2023

  3. [3]

    Esrgan: Enhanced super-resolution generative adversarial networks,

    Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy, “Esrgan: Enhanced super-resolution generative adversarial networks,” in Proceedings of the European Conference on Computer Vision Workshops (ECCVW) , 2018

    work page 2018

  4. [4]

    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,” in Proceedings of the IEEE International Conference on Computer Vision Workshops (ICCVW), 2021

    work page 2021

  5. [5]

    Denoising diffusion probabilistic models,

    Jonathan Ho, Ajay Jain, and Pieter Abbeel, “Denoising diffusion probabilistic models,” in Advances in Neural Information Processing Systems, 2020, pp. 1, 2, 3

    work page 2020

  6. [6]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, et al., “High-resolution image synthesis with latent diffusion models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 10684–10695

    work page 2022

  7. [7]

    Adding conditional control to text-to-image diffusion models,

    L. Zhang, A. Rao, and M. Agrawala, “Adding conditional control to text-to-image diffusion models,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 3836–3847

    2023

  8. [8]

    Image superresolution via iterative refinement,

    Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi, “Image superresolution via iterative refinement,” TPAMI, 2022

    work page 2022

Show all 39 references
  1. [9]

    Repaint: Inpainting using denoising diffusion probabilistic models,

    Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool, “Repaint: Inpainting using denoising diffusion probabilistic models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 11461–11471

    2022

  2. [10]

    Deblurring via stochastic refinement,

    Jay Whang, Mauricio Delbracio, Hossein Talebi, Chitwan Saharia, Alexandros G Dimakis, and Peyman Milanfar, “Deblurring via stochastic refinement,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 16293–16303

    2022

  3. [11]

    Ilvr: Conditioning method for denoising diffusion probabilistic models,

    J. Choi, S. Kim, Y . Jeong, Y . Gwon, and S. Yoon, “Ilvr: Conditioning method for denoising diffusion probabilistic models,” arXiv preprint arXiv:2108.02938, 2021

    2021 arXiv

  4. [12]

    Generative diffusion prior for unified image restoration and enhancement,

    B. Fei, Z. Lyu, L. Pan, et al., “Generative diffusion prior for unified image restoration and enhancement,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 9935–9946

    2023

  5. [13]

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

    Jian Wang, Ziqi Yue, Shuyang Zhou, et al., “Exploiting diffusion prior for real-world image super-resolution,” International Journal of Computer Vision, pp. 1–21, 2024

    2024

  6. [14]

    Denoising diffusion implicit models,

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

    2010 arXiv

  7. [15]

    Image super-resolution using deep convolutional networks,

    Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang, “Image super-resolution using deep convolutional networks,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 38, no. 2, pp. 295– 307, 2015

    2015

  8. [16]

    Compression artifacts reduction by a deep convolutional network,

    Chao Dong, Yi Deng, and Chen Change Loy, “Compression artifacts reduction by a deep convolutional network,” in Proceedings of the IEEE International Conference on Computer Vision (ICCV) , 2015, pp. 576– 584

    2015

  9. [17]

    Attention is all you need,

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin, “Attention is all you need,” in Advances in Neural Information Processing Systems (NeurIPS), 2017, pp. 5998–6008

    2017

  10. [18]

    Pre-trained image processing transformer,

    Haochen Chen, Yixuan Wang, Tian Guo, Chao Dong, Ziwei Liu, and Ping Luo, “Pre-trained image processing transformer,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 12299–12310

    2021

  11. [19]

    Restormer: Efficient transformer for high-resolution image restoration,

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang, “Restormer: Efficient transformer for high-resolution image restoration,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  12. [20]

    Gen- erative adversarial nets,

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio, “Gen- erative adversarial nets,” in Advances in Neural Information Processing Systems (NeurIPS), 2014

    2014

  13. [21]

    Diffusart: Enhancing line art colorization with conditional diffusion models,

    H ´ector Carrillo, Marc Cl ´ement, Aude Bugeau, et al., “Diffusart: Enhancing line art colorization with conditional diffusion models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 3486–3490

    2023

  14. [22]

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

    Yun Wang, Wei Yang, Xin Chen, et al., “Sinsr: Diffusion-based image super-resolution in a single step,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 25796–25805

    2024

  15. [23]

    Diffir: Efficient diffusion model for image restoration,

    B. Xia, Y . Zhang, S. Wang, et al., “Diffir: Efficient diffusion model for image restoration,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 13095–13105

    2023

  16. [24]

    Multiscale structure guided diffusion for image deblurring,

    Ming Ren, Matan Delbracio, Hamid Talebi, et al., “Multiscale structure guided diffusion for image deblurring,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 10721– 10733

    2023

  17. [25]

    Learning transferable visual models from natural language supervision,

    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 supervision,” in International conference on machine learning....

    2021

  18. [26]

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

    J. Wang, Z. Yue, S. Zhou, et al., “Exploiting diffusion prior for real-world image super-resolution,” International Journal of Computer Vision, pp. 1–21, 2024

    2024

  19. [27]

    To- wards robust blind face restoration with codebook lookup transformer,

    Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy, “To- wards robust blind face restoration with codebook lookup transformer,” Advances in Neural Information Processing Systems, vol. 35, pp. 30599– 30611, 2022

    2022

  20. [28]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R Zhang, P Isola, AA Efros, E Shechtman, and O Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018

    2018

  21. [29]

    Exploring clip for assessing the look and feel of images,

    J Wang, KC Chan, and CC Loy, “Exploring clip for assessing the look and feel of images,” in Proceedings of the AAAI Conference on Artificial Intelligence, 2023

    2023

  22. [30]

    Musiq: Multi- scale image quality transformer,

    J Ke, Q Wang, Y Wang, P Milanfar, and F Yang, “Musiq: Multi- scale image quality transformer,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , 2021

    2021

  23. [31]

    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, 6

    2017

  24. [32]

    jantic/DeOldify: A Deep Learning based project for colorizing and restoring old images (and video!),

    Anctic et al., “jantic/DeOldify: A Deep Learning based project for colorizing and restoring old images (and video!),” https://github.com/ jantic/DeOldify, 2021

    2021

  25. [33]

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

    Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang, “Toward real-world single image super-resolution: A new benchmark and a new model,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , 2019

    2019

  26. [34]

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

    Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin, “Component divide-and-conquer for real-world image super-resolution,” in Proceedings of the European Conference on Computer Vision (ECCV) , 2020

    2020

  27. [35]

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

    X Wang, L Xie, C Dong, and Y Shan, “Real-esrgan: Training real- world blind super-resolution with pure synthetic data,” in Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCV-W), 2021

    2021

  28. [36]

    Cdformer: When degradation prediction embraces diffusion model for blind image super-resolution,

    Q. Liu, C. Zhuang, P. Gao, et al., “Cdformer: When degradation prediction embraces diffusion model for blind image super-resolution,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 7455–7464

    2024

  29. [37]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models,

    J. Li, D. Li, S. Savarese, et al., “Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models,” in International Conference on Machine Learning. 2023, pp. 19730–19742, PMLR

    2023

  30. [38]

    Unicolor: A unified framework for multi-modal coloriza- tion with transformer,

    Huang et al., “Unicolor: A unified framework for multi-modal coloriza- tion with transformer,” ACM Transactions on Graphics (TOG) , vol. 41, no. 6, pp. 1–16, 2022

    2022

  31. [39]

    L-code: Language-based colorization using color-object decoupled conditions,

    S Weng, H Wu, Z Chang, et al., “L-code: Language-based colorization using color-object decoupled conditions,” in Proceedings of the AAAI Conference on Artificial Intelligence , 2022, vol. 36, pp. 2677–2684

    2022

Pith tools

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