Pith. sign in

REVIEW 3 major objections 6 minor 42 references

UniLDiff: Unlocking the Power of Diffusion Priors for All-in-One Image Restoration

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read UniLDiff claims that a single diffusion model, extended with degradation-aware feature fusion and detail-aware expert modules, achieves state-of-the-art perceptual quality across multiple and mixed image restoration tasks without explicit…

desk verdict Solid incremental engineering with strong empirical coverage, but the DAEM mechanism claim is confounded by decoder fine-tuning in the ablation, and missing code/error bars make the modest gains hard to evaluate. read the letter →

arxiv 2507.23685 v2 pith:EPJBCFVI submitted 2025-07-31 cs.CV

Zihan Cheng , Liangtai Zhou , Dian Chen , Ni Tang , Xiaotong Luo , Yanyun Qu This is my paper
classification cs.CV
keywords all-in-oneimagerestorationlatentdiffusionmodelspriorsdegradation-awarefeaturefusiondetail-awareexpertmodulemixture-of-expertsperceptualqualitycompositedegradation
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 presents UniLDiff, a single diffusion-based model that aims to handle many image restoration tasks—denoising, deraining, dehazing, deblurring, and low-light enhancement—and their combinations, without being told which degradation is present. It claims that two additions to a pre-trained latent diffusion model are enough to unlock this generality: a Degradation-Aware Feature Fusion (DAFF) that feeds the corrupted image's features into every denoising step through adaptive, timestep-dependent gating, and a Detail-Aware Expert Module (DAEM) that routes decoder features through specialized branches to recover high-frequency texture. If the claims hold, unified restoration no longer depends on task labels or pre-classifying the degradation, which matters because real-world images often suffer from compound, spatially uneven flaws. On standard benchmarks the method reports the best perceptual-quality scores across multi-task and mixed-degradation settings while remaining competitive in pixel-level fidelity.

What carries the argument

The machinery is two interlocking modules retrofitted onto a frozen SDXL VAE–UNet pipeline. DAFF (Degradation-Aware Feature Fusion) takes the low-quality image's multi-scale features and the noisy clean latent at each timestep, runs them through decoupled double-stream attention plus a single-stream fusion block, and uses the timestep $t$ as a conditioning signal to modulate how strongly LQ cues influence the current denoising step. DAEM (Detail-Aware Expert Module) sits in the VAE decoder: a router selects one expert (a lightweight NAFBlock with a particular receptive field) per spatial region, multiplies the expert output by a shared transposed self-attention branch, and fuses in skip-connected encoder features that carried high-resolution cues through VAE compression. Together they turn the diffusion model's generative prior into a degradation-adaptive restorer: DAFF decides what kind of corruption is present, and DAEM decides where to spend detail-recovery effort.

What would settle it

Take a clean high-quality image, pass it through the frozen SDXL VAE encoder–decoder, and measure the high-frequency energy or edge-texture similarity (for example, FFT band error or LPIPS on band-passed images) between the reconstruction and the original. If the VAE's own reconstruction already loses most high-frequency content, then run UniLDiff on a corrupted version of the same image and compute the same high-frequency fidelity against the original. The detail-recovery claim stands only if the restored image's high-frequency error is substantially smaller than the VAE's own reconstruction error, not merely better than other restorers.

Watch

Extended reading notes

Core claim

UniLDiff is an LDM-based unified image restoration framework built on the Stable Diffusion XL backbone. The central claim is that injecting low-quality (LQ) image features into the denoising trajectory through the DAFF module—a cascaded double-stream and single-stream attention fusion with timestep-aware modulation—lets the diffusion prior implicitly adapt to arbitrary and compound degradations without predefined degradation types. A second claim is that the DAEM, a mixture-of-experts decoder module with skip-connected encoder features and a global transposed self-attention branch, recovers high-frequency details lost in VAE compression and improves texture and structure fidelity. The paper reports state-of-the-art performance on perceptual metrics (LPIPS, DISTS, CLIPIQA, MUSIQ, MANIQA) across the three-task and five-task settings and the best average PSNR/SSIM on the CDD11 composite-degradation benchmark (29.35 dB / 0.886).

Load-bearing premise

The framework assumes that the frozen VAE decoder, aided by DAEM and the trainable LQ encoder's skip-connected features, can actually reconstruct the fine textures that the VAE encoder's high compression discarded; if that bottleneck is too severe, the added details would be plausible hallucinations rather than restored truth.

Editorial extensions

If this is right

  • A single trained model restores five different degradation types and their composites with one set of weights, so deployment shifts from maintaining per-task models to one backbone.
  • Because DAFF conditions on the image itself rather than a label or global prompt, the method extends to spatially heterogeneous and compound degradations without retraining or prompt engineering.
  • DAEM's mixture-of-experts design with a load-balancing loss prevents expert collapse, meaning detail-recovery capacity scales with the number and width of experts rather than with distillation tricks.
  • Perceptual-quality gains on LPIPS, DISTS, and no-reference metrics come without sacrificing competitive PSNR/SSIM on composite benchmarks, so the method claims both fidelity and realism.

Reading between the lines

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

  • The DAFF design implies that explicit degradation classification may be subsumed by per-step attention-based fusion; a testable extension is whether the same module handles unseen degradation types such as super-resolution or compression artifacts without architectural change.
  • The paper's own admission that the VAE encoder discards high-frequency signals suggests recovered details may partly be hallucinated rather than retrieved; a natural next experiment measures high-frequency fidelity against ground-truth edge spectra rather than aggregate perceptual metrics.
  • Since DAEM activates one expert per region, the effective receptive field per patch is a discrete choice; an extension would let the router pick a continuous blend of experts to smooth boundaries between differently degraded regions.
  • The method still uses an optional task prompt alongside DAFF, but the ablations show DAFF drives the main gain, so a fully label-free variant without any text prompt is within reach.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. UniLDiff proposes a unified latent diffusion framework for all-in-one image restoration (AiOIR). The method combines a Degradation-Aware Feature Fusion (DAFF) module that injects low-quality features into each denoising step, and a Detail-Aware Expert Module (DAEM) that routes decoder features through a mixture-of-experts to enhance high-frequency details. Training is done in two stages: first DAFF is trained with the diffusion UNet, then the VAE decoder is fine-tuned together with DAEM. Experiments cover three-task and five-task universal restoration, plus the CDD11 composite-degradation benchmark, with extensive comparisons against recent CNN-, Transformer-, and diffusion-based baselines. The main claims are state-of-the-art perceptual quality on the multi-task settings and competitive or leading PSNR/SSIM on composite degradations.

Significance. If the claims hold, UniLDiff would be a useful contribution to AiOIR: it demonstrates that degradation-aware feature fusion inside a diffusion UNet can handle multiple and compound degradations without predefined degradation labels, and that a decoder-side expert module can improve perceptual fidelity. The paper ships an unusually broad set of experiments, including no-reference perceptual metrics, five-task joint training, and the CDD11 composite benchmark, which strengthens the empirical picture. The authors also correctly identify a known weakness of latent diffusion models—VAE compression discarding high-frequency detail—and attempt to address it explicitly. However, the central contribution, DAEM, is not yet validated by the ablation because the ablation confounds DAEM with VAE decoder fine-tuning; this must be resolved before the mechanism claim is credible.

major comments (3)
  1. [Method (Stage 2) and Table 5] The ablation in Table 5 does not isolate the contribution of DAEM. Stage 2 is described as 'we fine-tune the V AE decoder together with the DAEM,' but the baseline rows without DAEM (e.g., 'DAFF + Task Prompt' with PSNR 27.36) are not described as undergoing any decoder fine-tuning. If those baselines use the original frozen SDXL decoder, then the 2.91 dB PSNR gain and the MUSIQ gain from adding DAEM conflate two independent changes: decoder adaptation to the restoration task and the DAEM routing mechanism itself. A decoder fine-tuned with Lrecon and Lssim on restoration targets could plausibly account for a large part of this improvement regardless of expert routing. Please add a control that fine-tunes the decoder alone (without DAEM) while keeping all other components fixed, or an ablation that adds DAEM while keeping the decoder frozen. Without such a control, the paper's claim that DAEM specifically enhances texture and fine-structure recovery is unsupported.
  2. [Experiment (CDD11) and Tables 7] The average PSNR advantage over MoCE-IR-S on CDD11 is 0.30 dB (29.35 vs 29.05) and the average SSIM advantage is 0.005. No error bars, confidence intervals, or significance tests are reported anywhere in the paper. Given that this margin is small relative to typical run-to-run variation in image restoration training, the statement that the method 'consistently delivers strong performance' and 'outperforms' the baseline in distortion metrics is stronger than the evidence supports. I request per-seed standard deviations or statistical testing for the aggregate CDD11 results, and ideally for the key perceptual-metric tables as well.
  3. [Appendix and Method (Unified Training Strategy)] Two promised supporting details are missing. First, the Method section states that 'Details of its formulation are provided in the Appendix' for the auxiliary load-balancing loss Laux, but the appendix contains no such formulation. Second, the Experiment section states that 'More detailed results including runtime and parameter statistics are provided in Appendix,' but the appendix does not report runtime or parameter counts. These are asserted rather than delivered pieces of evidence; please add them or revise the statements.
minor comments (6)
  1. [Method, Eq. (9)] The norm in Eq. (9) is written as double vertical bars without a subscript; since the text calls it the standard diffusion noise objective, please specify whether this is an L1 or L2 norm and use the corresponding notation.
  2. [Method, Eq. (6)] Eq. (6) uses DDPM-style notation with alpha_t, sigma_t, and z, but the relationship between sigma_t and the alpha schedule is not defined, and z appears without a coefficient. Please clarify the sampling formula and how sigma_t is computed.
  3. [Method, DAEM and DAFF] Several key hyperparameters are stated only as 'empirically set': the number of experts, the router noise scale sigma, the activation function phi, and the loss weights lambda_1 and lambda_2. Please provide the values used and, ideally, a small sensitivity study for at least lambda_1 and lambda_2.
  4. [Tables 3 and 6] Table 3 uses red/blue for best/second-best while Table 6 uses bold for best; the inconsistent formatting conventions make the tables harder to read. Please unify the highlighting convention.
  5. [Abstract and Table 1] The abstract claims the method 'consistently achieves state-of-the-art performance,' but in Table 1 the PSNR and SSIM are lower than several non-diffusion baselines. Please qualify the claim to specify that the state-of-the-art results are for perceptual metrics, or discuss the distortion-perception tradeoff explicitly.
  6. [Figure 5 caption] The phrase 'in some cases surpassing the visual quality of the ground truth' is subjective and potentially confusing; it would be more precise to say 'restoring details that are not clearly present in the ground-truth image' or simply describe measurable improvements.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the paper's claims are supported by held-out benchmark evaluations rather than by construction or self-citation.

full rationale

UniLDiff is an empirical, supervised image-restoration paper. The claimed 'predictions' are test-set results on standard benchmarks (BSD68, Rain100L, RESIDE-SOTS, GoPro, LOL-v1, CDD11) compared against existing published methods; they are not fitted quantities derived from the model's own definitions. The training objectives in Eq. (9) and Eq. (10) are conventional diffusion noise-estimation and reconstruction losses, and the reported metrics are external evaluations of the resulting model. The DAFF and DAEM modules are trained network components whose contributions are assessed by ablations, not by identities or by construction. There is no load-bearing self-citation chain: references to FLUX, SDXL, NAFBlock, and Restormer are external architectural inspirations or pretrained components, and no 'uniqueness theorem' or prior result by the same authors is invoked to force the design. The only notable methodological concern is that the Table 5 ablation of DAEM may be confounded with Stage-2 VAE-decoder fine-tuning, since the baseline row without DAEM is not described as undergoing decoder fine-tuning; however, this is an experimental attribution issue rather than a circular derivation, and it does not make any reported result equivalent to the paper's inputs by construction.

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

The paper introduces no new physical or conceptual entities. Its contribution is an architecture built from existing components (SDXL, NAFBlock, Restormer attention, FLUX-style fusion, MoE). The central claim sells on hand-picked hyperparameters and the assumption that the benchmark tasks capture real-world degradation.

free parameters (3)
  • k (number of active experts) = 1 (set empirically)
    Set empirically in DAEM router (Eq. 7); the paper says 'we empirically set k=1'.
  • lambda_1, lambda_2 (loss weights) = not reported
    Used in Eq. 10 for L_stage-2; values are chosen by hand but not given.
  • sigma (router noise scale) = not reported
    Gaussian noise in the routing function (Eq. 7); scale not specified.
assumptions (3)
  • domain assumption Pre-trained SDXL VAE, UNet, and text encoder provide an appropriate prior for image restoration.
    Invoked in Method: 'we adopt Stable Diffusion XL as our underlying latent diffusion model, leveraging its pre-trained components to initialize the encoder, decoder, and denoising UNet.'
  • domain assumption The standard denoising objective (Eq. 9) and the Stage-2 composite loss (Eq. 10) are sufficient to train the restoration model.
    Equations 9 and 10 define the training losses; the paper does not ablate alternative objectives.
  • domain assumption Evaluation on BSD68, Rain100L, RESIDE-SOTS, GoPro, LOL-v1, and CDD11 is representative of real-world restoration performance.
    Used in Experiments section; the transfer from synthetic benchmarks to in-the-wild images is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UniLDiff: Unlocking the Power of Diffusion Priors for All-in-One Image Restoration." pith.science (2026). https://pith.science/paper/EPJBCFVI

@misc{pith2026250723685,
  author       = {Pith},
  title        = {Pith review of: UniLDiff: Unlocking the Power of Diffusion Priors for All-in-One Image Restoration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EPJBCFVI}},
  note         = {Machine review of arXiv:2507.23685}
}
read the original abstract

All-in-One Image Restoration (AiOIR) has emerged as a promising yet challenging research direction. To address the core challenges of diverse degradation modeling and detail preservation, we propose UniLDiff, a unified framework enhanced with degradation- and detail-aware mechanisms, unlocking the power of diffusion priors for robust image restoration. Specifically, we introduce a Degradation-Aware Feature Fusion (DAFF) to dynamically inject low-quality features into each denoising step via decoupled fusion and adaptive modulation, enabling implicit modeling of diverse and compound degradations. Furthermore, we design a Detail-Aware Expert Module (DAEM) in the decoder to enhance texture and fine-structure recovery through expert routing. Extensive experiments across multi-task and mixed degradation settings demonstrate that our method consistently achieves state-of-the-art performance, highlighting the practical potential of diffusion priors for unified image restoration. Our code will be released.

Figures

Figures reproduced from arXiv: 2507.23685 by the authors.

Figure 1
Figure 1. MUSIQ-based comparison across five restoration [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overall architecture of the proposed UniLDiff framework. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Architecture of the proposed Degradation-Aware [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Architecture of the DAEM, which uses experts [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Visual results. In dehazing, our approach produces more natural color restoration and sharper details, in some cases surpassing the visual quality of the ground truth. For low-light enhancement, it better preserves structural edges and surface textures while avoiding o…
Figure 6
Figure 6. Figure 6: Effectiveness of the DAEM. on local context. As shown in [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Qualitative comparison on five single degradation types: deraining, denoising, dehazing, low-light enhancement, and [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Qualitative comparison on the CDD11 dataset. Our method shows strong generalization and high-quality restoration [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 7 canonical work pages

  1. [1]

    Ai, Y.; Huang, H.; Zhou, X.; Wang, J.; and He, R. 2024. Multimodal prompt perceiver: Empower adaptiveness generalizability and fidelity for all-in-one image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 25432--25444

    2024

  2. [2]

    Arbelaez, P.; Maire, M.; Fowlkes, C.; and Malik, J. 2010. Contour detection and hierarchical image segmentation. IEEE transactions on pattern analysis and machine intelligence, 33(5): 898--916

    2010

  3. [3]

    Chen, J.; Pan, J.; and Dong, J. 2025. Faithdiff: Unleashing diffusion priors for faithful image super-resolution. In Proceedings of the Computer Vision and Pattern Recognition Conference, 28188--28197

    work page 2025

  4. [4]

    Chen, L.; Chu, X.; Zhang, X.; and Sun, J. 2022 a . Simple baselines for image restoration. In European conference on computer vision, 17--33. Springer

    work page 2022

  5. [5]

    Chen, W.-T.; Huang, Z.-K.; Tsai, C.-C.; Yang, H.-H.; Ding, J.-J.; and Kuo, S.-Y. 2022 b . Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 17653--17662

    work page 2022

  6. [6]

    W.; Khan, S.; Knoll, A.; Shah, M.; and Khan, F

    Cui, Y.; Zamir, S. W.; Khan, S.; Knoll, A.; Shah, M.; and Khan, F. S. 2024. Adair: Adaptive all-in-one image restoration via frequency mining and modulation. arXiv preprint arXiv:2403.14614

    arXiv 2024

  7. [7]

    Dai, X.; Hou, J.; Ma, C.-Y.; Tsai, S.; Wang, J.; Wang, R.; Zhang, P.; Vandenhende, S.; Wang, X.; Dubey, A.; et al. 2023. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807

    arXiv 2023

  8. [8]

    Ding, K.; Ma, K.; Wang, S.; and Simoncelli, E. P. 2020. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5): 2567--2581

    2020

Show all 42 references
  1. [9]

    Duan, H.; Min, X.; Wu, S.; Shen, W.; and Zhai, G. 2024. Uniprocessor: a text-induced unified low-level image processor. In European Conference on Computer Vision, 180--199. Springer

    2024

  2. [10]

    W.; and He, S

    Guo, Y.; Gao, Y.; Lu, Y.; Zhu, H.; Liu, R. W.; and He, S. 2024. Onerestore: A universal restoration framework for composite degradation. In European conference on computer vision, 255--272. Springer

    2024

  3. [11]

    Jiang, J.; Zuo, Z.; Wu, G.; Jiang, K.; and Liu, X. 2024 a . A survey on all-in-one image restoration: Taxonomy, evaluation and future trends. arXiv preprint arXiv:2410.15067

    2024 arXiv

  4. [12]

    Jiang, Y.; Zhang, Z.; Xue, T.; and Gu, J. 2024 b . Autodir: Automatic all-in-one image restoration with latent diffusion. In European Conference on Computer Vision, 340--359. Springer

    2024

  5. [13]

    Ke, J.; Wang, Q.; Wang, Y.; Milanfar, P.; and Yang, F. 2021. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, 5148--5157

    2021

  6. [14]

    P.; Welling, M.; et al

    Kingma, D. P.; Welling, M.; et al. 2013. Auto-encoding variational bayes

    2013

  7. [15]

    Labs, B. F.; Batifol, S.; Blattmann, A.; Boesel, F.; Consul, S.; Diagne, C.; Dockhorn, T.; English, J.; English, Z.; Esser, P.; Kulal, S.; Lacey, K.; Levi, Y.; Li, C.; Lorenz, D.; Müller, J.; Podell, D.; Rombach, R.; Saini, H.; Sauer, A.; and Smith, L. 2025. FLUX.1 Kontext: Fl...

    2025 arXiv

  8. [16]

    Li, B.; Liu, X.; Hu, P.; Wu, Z.; Lv, J.; and Peng, X. 2022. All-in-one image restoration for unknown corruption. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 17452--17462

    2022

  9. [17]

    Li, B.; Ren, W.; Fu, D.; Tao, D.; Feng, D.; Zeng, W.; and Wang, Z. 2018. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1): 492--505

    2018

  10. [18]

    o lund, J.; and Sch \

    Luo, Z.; Gustafsson, F. K.; Zhao, Z.; Sj \"o lund, J.; and Sch \"o n, T. B. 2023. Controlling vision-language models for universal image restoration. arXiv preprint arXiv:2310.01018, 3(8)

    2023 arXiv

  11. [19]

    Ma, K.; Duanmu, Z.; Wu, Q.; Wang, Z.; Yong, H.; Li, H.; and Zhang, L. 2016. Waterloo exploration database: New challenges for image quality assessment models. IEEE Transactions on Image Processing, 26(2): 1004--1016

    2016

  12. [20]

    Martin, D.; Fowlkes, C.; Tal, D.; and Malik, J. 2001. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings eighth IEEE international conference on computer vision. ICCV 2001, ...

    2001

  13. [21]

    Nah, S.; Hyun Kim, T.; and Mu Lee, K. 2017. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, 3883--3891

    2017

  14. [22]

    Podell, D.; English, Z.; Lacey, K.; Blattmann, A.; Dockhorn, T.; M \"u ller, J.; Penna, J.; and Rombach, R. 2023. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952

    2023 arXiv

  15. [23]

    W.; Khan, S

    Potlapalli, V.; Zamir, S. W.; Khan, S. H.; and Shahbaz Khan, F. 2023. Promptir: Prompting for all-in-one image restoration. Advances in Neural Information Processing Systems, 36: 71275--71293

    2023

  16. [24]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, 8748--8763. PmLR

    2021

  17. [25]

    Riquelme, C.; Puigcerver, J.; Mustafa, B.; Neumann, M.; Jenatton, R.; Susano Pinto, A.; Keysers, D.; and Houlsby, N. 2021. Scaling vision with sparse mixture of experts. Advances in Neural Information Processing Systems, 34: 8583--8595

    2021

  18. [26]

    Tian, X.; Liao, X.; Liu, X.; Li, M.; and Ren, C. 2025. Degradation-Aware Feature Perturbation for All-in-One Image Restoration. In Proceedings of the Computer Vision and Pattern Recognition Conference, 28165--28175

    2025

  19. [27]

    Valanarasu, J. M. J.; Yasarla, R.; and Patel, V. M. 2022. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2353--2363

    2022

  20. [28]

    C.; and Loy, C

    Wang, J.; Chan, K. C.; and Loy, C. C. 2023. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI conference on artificial intelligence, volume 37, 2555--2563

    2023

  21. [29]

    C.; Sheikh, H

    Wang, Z.; Bovik, A. C.; Sheikh, H. R.; and Simoncelli, E. P. 2004. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4): 600--612

    2004

  22. [30]

    Wei, C.; Wang, W.; Yang, W.; and Liu, J. 2018. Deep retinex decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560

    2018 arXiv

  23. [31]

    Yan, Q.; Jiang, A.; Chen, K.; Peng, L.; Yi, Q.; and Zhang, C. 2025. Textual prompt guided image restoration. Engineering Applications of Artificial Intelligence, 155: 110981

    2025

  24. [32]

    Yang, S.; Wu, T.; Shi, S.; Lao, S.; Gong, Y.; Cao, M.; Wang, J.; and Yang, Y. 2022. Maniqa: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 1191--1200

    2022

  25. [33]

    T.; Feng, J.; Liu, J.; Guo, Z.; and Yan, S

    Yang, W.; Tan, R. T.; Feng, J.; Liu, J.; Guo, Z.; and Yan, S. 2017. Deep joint rain detection and removal from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, 1357--1366

    2017

  26. [34]

    P.; Zhang, Y.; and Timofte, R

    Zamfir, E.; Wu, Z.; Mehta, N.; Tan, Y.; Paudel, D. P.; Zhang, Y.; and Timofte, R. 2025. Complexity experts are task-discriminative learners for any image restoration. In Proceedings of the Computer Vision and Pattern Recognition Conference, 12753--12763

    2025

  27. [35]

    W.; Arora, A.; Khan, S.; Hayat, M.; Khan, F

    Zamir, S. W.; Arora, A.; Khan, S.; Hayat, M.; Khan, F. S.; and Yang, M.-H. 2022. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 5728--5739

    2022

  28. [36]

    Zeng, H.; Wang, X.; Chen, Y.; Su, J.; and Liu, J. 2025. Vision-Language Gradient Descent-driven All-in-One Deep Unfolding Networks. In Proceedings of the Computer Vision and Pattern Recognition Conference, 7524--7533

    2025

  29. [37]

    Zhang, L.; Zhang, L.; and Bovik, A. C. 2015. A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing, 24(8): 2579--2591

    2015

  30. [38]

    A.; Shechtman, E.; and Wang, O

    Zhang, R.; Isola, P.; Efros, A. A.; Shechtman, E.; and Wang, O. 2018. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, 586--595

    2018

  31. [39]

    Zheng, D.; Wu, X.-M.; Yang, S.; Zhang, J.; Hu, J.-F.; and Zheng, W.-S. 2024. Selective hourglass mapping for universal image restoration based on diffusion model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 25445--25455

    2024

  32. [40]

    Zhu, Z.; Feng, X.; Chen, D.; Bao, J.; Wang, L.; Chen, Y.; Yuan, L.; and Hua, G. 2023. Designing a better asymmetric vqgan for stablediffusion. arXiv preprint arXiv:2306.04632

    2023 arXiv

  33. [41]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  34. [42]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.