Pith. sign in

REVIEW 4 major objections 5 minor 67 references

Exploring Scalable Unified Modeling for General Low-Level Vision

T0 review · 4 major / 5 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 GenLV is a single prompt-conditioned model that unifies more than 100 low-level vision tasks and scales with model capacity and task diversity.

desk verdict Useful scaling study with a flawed evaluation protocol: the specialist-comparison claim is contradicted by the paper's own tables, and the task-diversity claim lacks a controlled sweep. read the letter →

arxiv 2507.14801 v1 pith:LC2G4MMV submitted 2025-07-20 cs.CV

Xiangyu Chen , Kaiwen Zhu , Yuandong Pu , Shuo Cao , Xiaohui Li , Wenlong Zhang , Yihao Liu , Yu Qiao
show 2 more authors
Jiantao Zhou Chao Dong
This is my paper · ORCID
classification cs.CV
keywords low-levelvisionvisualpromptingmulti-tasklearningimagerestorationenhancementstylizationfeatureextractionfoundationmodel
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 sets out to show that low-level vision need not be split into hundreds of task-specific networks: one model, guided by visual prompt pairs, can handle restoration, enhancement, stylization, and feature extraction together. It introduces GenLV, trained on a 101-task benchmark, and argues that expanding the task set improves generalization, especially for tasks with little data, while enlarging the model yields consistent gains across most task types. The authors further claim that with only a handful of examples or brief fine-tuning, the same model can adapt to unseen real-world tasks and match or beat dedicated specialists. A sympathetic reader would take the core contribution as evidence that visual prompting plus a strong reconstruction backbone is a viable path toward a general low-level vision foundation model.

What carries the argument

The load-bearing mechanism is the Visual task Prompt-based Image Processing (VPIP) framework, built from an end-to-end image processing backbone, a visual prompt encoder, and a Prompt Cross-Attention Block (PCAB). The prompt pair $[P_{\Omega_S}, P_{\Omega_T}]$ is encoded into two spatial feature maps, and PCAB mixes them into the input features using cross-attention with $Q = W_Q(z)$, $K = W_K(z^P_{\Omega_S})$, $V = W_V(z^P_{\Omega_T})$, followed by scaled dot-product attention. This lets the model condition the same weights on the difference between source and target domains, decoupling task guidance from the backbone architecture. The backbone is X-Restormer, a U-shaped network with channel-wise and spatial self-attention blocks, chosen for pixel-level reconstruction fidelity; this combination avoids the ViT-and-MAE constraints the paper attributes to earlier generalist models such as Painter and PromptGIP.

What would settle it

Train a dedicated specialist model for each of the 101 tasks in GLV-Bench using the same Restormer backbone, then compare per-task PSNR, SSIM, LPIPS, and MAE against GenLV-Huge. If the specialists beat GenLV-Huge on the majority of tasks where the paper claims parity or superiority, the reference-model proxy is too weak to support the generality claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that a single prompt-conditioned network can unify heterogeneous low-level vision tasks. GenLV takes an input image together with a visual prompt pair, one example from the source domain and one from the target domain, and infers the intended transformation from those images alone. Trained on GLV-Bench, a benchmark of 101 tasks spanning restoration, enhancement, stylization, and feature extraction, the model performs competitively with multi-task reference models on most tasks and sometimes exceeds task-specific specialists, particularly where training data are scarce. The paper also claims that increasing model capacity produces substantial and consistent improvements across most task categories, and that few-shot fine-tuning with as few as five sample pairs turns GenLV into a specialist that matches or surpasses published task-specific models on real denoising, deblurring, and super-resolution.

Load-bearing premise

The evaluation assumes that a single multi-task Restormer model trained on a hand-defined group of related tasks is a reliable stand-in for a dedicated specialist on every task in that group, so that beating the reference model counts as evidence of true generality rather than of a weak baseline.

Editorial extensions

If this is right

  • A single trained model can serve dozens of low-level tasks without per-task architectures or per-task training runs.
  • Adding more training tasks can improve performance on data-scarce tasks, so joint training acts as a form of knowledge transfer.
  • Scaling model capacity yields consistent gains on restoration, enhancement, and stylization, with the main exception being semantic-heavy feature extraction tasks like depth and normal estimation.
  • The visual-prompt design is stable across different prompt images, with standard deviations around or below 0.1 dB on most tasks, making prompt choice a reliable control signal.
  • Fine-tuning the pre-trained model with very few samples can match or beat dedicated specialists, suggesting practical use as a low-level vision foundation model.

Reading between the lines

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

  • An implication the authors leave implicit is that the reference-model evaluation protocol could be strengthened: training a dedicated specialist for every one of the 101 tasks would give a harder baseline, and the current group-wise surrogate is demonstrably unreliable for heterogeneous groups such as dehazing, where the reference model collapses to 18.94 dB while a specialist reaches 31.90 dB.
  • If the scaling trend continues, larger versions of GenLV may narrow the remaining gap on semantic feature extraction tasks, which currently appear limited by task difficulty rather than by model capacity alone.
  • A natural testable extension would hold the task set fixed and vary the task grouping, asking whether the observed positive transfer is predictable from pairwise task similarity; that would turn the benchmark into a tool for studying task synergy and interference.
  • The authors' visual-prompt interaction could be combined with text instructions to make the model more user-friendly, a direction they acknowledge only as a limitation rather than exploring.
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 / 5 minor

Summary. The paper proposes VPIP, a visual-prompt-conditioned framework for unified low-level vision, and instantiates it as GenLV, an X-Restormer-based model guided by input-target image pairs. It first validates GenLV on 30 tasks against Painter and PromptGIP, then introduces a 101-task benchmark, GLV-Bench, and trains three model variants (Base, Large, Huge) to study scaling. Evaluation is carried out against per-group Restormer reference models and a small set of specialist models, followed by zero-shot, few-shot, and task-specific fine-tuning experiments. The central claims are that the framework unifies restoration, enhancement, stylization, and feature extraction, that enlarging capacity yields consistent gains, that increasing task diversity improves generalization, and that the model has foundation-model potential.

Significance. If fully supported, the paper would be a substantial contribution: GLV-Bench is a large and heterogeneous public benchmark, the VPIP design decouples prompt conditioning from the backbone in a sensible way, the prompt-robustness analysis in Table V is strong, and the fine-tuning results in Table XI are encouraging positive evidence. The paper also reports honest negative results, such as the poor feature-extraction scaling and explicit failure cases. However, the current evidence does not support all headline claims: the reference-model protocol is demonstrably unreliable for several task groups, the statement that GenLV-Huge exceeds specialists in Groups 1-4 is contradicted by the paper's own tables, and the abstract's task-count scaling claim is not tested by a controlled experiment. The core architecture and benchmark are valuable, but the evaluation protocol and several textual claims need substantial revision.

major comments (4)
  1. [§IV-C.3, Tables VI-IX] The claim that GenLV-Huge 'in some cases, exceeds specialist model performance (e.g., Groups 1–4)' is contradicted by the paper's own specialist entries. Across the 20 specialist columns in Tables VI-IX, GenLV-Huge is worse than the specialist in 11 cases and better in 9. In Groups 1-4, it loses all six specialist comparisons: SRx4 (24.82 vs 25.94), CT SR (32.74 vs 32.84), Gaussian Deblurring (27.42 vs 29.05), Poisson Denoising (30.30 vs 31.37), Satellite Denoising (33.02 vs 33.60), and Inpainting (29.32 vs 31.25). The sentence should be corrected to name the specific tasks where GenLV-Huge actually surpasses the specialist, and the implied claim of general specialist-level performance should be removed or substantially softened.
  2. [§IV-C.1, Table VI] The evaluation protocol assumes that a single Restormer reference model trained per hand-defined group is a reliable proxy for task-specific specialist performance, but the paper's own data falsify this assumption for entire groups. For example, the reference model scores 18.94 dB on Dehazing versus a 31.90 dB specialist, 23.13 dB on Complex Deraining versus 26.66, 34.00 dB on Demoireing versus 39.86, and 29.31 dB on Highlight Removal versus 33.90. GenLV-Huge's largest margins over reference models are concentrated in exactly these groups, so those wins show only that the reference model is a weak baseline for those tasks, not that GenLV is generally competitive with specialists. The paper should either validate the reference-model proxy more rigorously, compare GenLV-Huge against the full set of specialist models, or clearly restrict the 'outperforms reference models' conclusion to tasks where the proxy is verified.
  3. [Abstract and §IV-C.3] The abstract's claim that 'increasing the number of training tasks enhances generalization, particularly for tasks with limited data' is not supported by a controlled task-count sweep. The 30-task model and the 101-task variants differ in prompt-encoder width (32 vs 64 channels), training epochs (30 vs 50), and task composition; there is no experiment that trains the same architecture on, say, 30, 50, and 101 tasks with all other settings held fixed. The conclusion in §IV-C.3 that 'when data is limited, more tasks can benefit models to generalize better' is inferred from comparing synthetic and fixed-dataset tasks, not from varying task count. A matched task-count ablation is needed, or the claim should be reformulated as a comparison of the 30-task and 101-task systems with all confounds acknowledged.
  4. [§IV-C.4 and Table XI] The fine-tuning experiments support the model's adaptability, but the 'cutting-edge specialists' used as baselines are not state of the art: CBDNet (2019), MPRNet (2020), Real-ESRNet (2021), and FunSR (2023) are now several generations old, and only four tasks are covered. The claim that fine-tuned GenLV 'matches or surpasses state-of-the-art specialist models' is therefore too strong. Either compare against current specialist methods on these tasks or rename the baselines 'representative specialist models' and qualify the conclusion.
minor comments (5)
  1. [Section V and Fig. 11] The model is referred to as 'GenLV100' in Section V and 'GenL V100' in the Fig. 11 caption, while the paper consistently uses 'GenLV' and 'GenLV-Huge' elsewhere; please unify the naming.
  2. [Tables VII and VIII] The group-index numbering is inconsistent: Group 15 is assigned to enhancement tasks (Local Laplacian Filtering, Multi-Scale TM, Bokeh Rendering) in Table VII and also to stylization tasks (Photographic, PencilDrawing, RTV) in Table VIII, while the first stylization block is labeled Group 16. This makes the claimed 18-group organization difficult to follow and should be corrected.
  3. [§IV-C.4] The text contains the typo 'GeinnLV is not only an effective general-purpose model'; this should read 'GenLV'.
  4. [§IV-C.1 and Tables VI-IX] For reproducibility, the paper should provide per-task details of the test sets, prompt-pair construction, and data sources used in GLV-Bench, as well as the training configuration (epochs, loss, optimizer) of the reference and specialist Restormer models; currently these details are only partially described.
  5. [§IV-B.1] The statement that GenLV† 'outperforms all baselines across most metrics' in Tables II-IV would benefit from explicit counts or error bars, since several entries are close and the tables report only point estimates.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: GenLV's benchmark claims are supported by measured comparisons against external baselines and specialists; the self-citations present are minor and non-essential.

full rationale

The paper's central claims—that VPIP enables a unified low-level vision model, that scaling model capacity helps, and that fine-tuning yields specialist-level performance—are supported by direct quantitative evaluations on fixed test sets (Common528, SIDD, RealBlur-J, DIV2K, AID, etc.) against external and self-constructed baselines, rather than by equations that reduce to their own inputs. No parameter is fitted to a subset and then renamed as a prediction; no 'uniqueness theorem' is imported to forbid alternatives; and the visual-prompt formulation, while inspired by prior work, is implemented and ablated within the paper (e.g., the ViT-VPIP vs. X-Restormer-VPIP comparisons in Table II). The self-citations to PromptGIP [10], the ACMMM2024 conference version [11], X-Restormer [34], and SEAL [56] are methodological or architectural references, but they are not load-bearing in the sense that the benchmark conclusions depend on an unverified self-cited result. The evaluation protocol using per-group Restormer reference models is explicitly described and even acknowledged in Sec. IV-C.3 to fail for some groups (e.g., Group 9 dehazing, Group 10 demoireing); this is a benchmarking-validity caveat, not a circular definition. The discrepancy between the prose claim that GenLV-Huge exceeds specialists in Groups 1-4 and the numbers in Table VI is a correctness/reporting concern, not a self-referential reduction. Overall, the derivation chain is empirically self-contained; the only reason the score is not zero is the presence of multiple minor author self-citations that are not load-bearing.

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

The central claims are empirical, not derived. The framework's validity depends on the prompt-representation assumption, the reference-model evaluation protocol, and synthetic target generation. The main hand-chosen design levers are the task grouping, the model capacity ladder, and the 256x256 training resolution. No new physical or theoretical entities are introduced.

free parameters (3)
  • Evaluation task grouping (18 groups) = n/a (hand-defined grouping)
    The 101 tasks are clustered into 18 groups to train Restormer reference models; the choice directly determines baseline strength, so measured GenLV-vs-reference gaps depend on it.
  • GenLV-Huge capacity configuration = channels [80, 160, 320, 640]; blocks [6, 8, 8, 12]; 211.11M params
    The scaling conclusion is measured across hand-chosen capacity steps; other capacity schedules could change the observed trend.
  • Training and prompt resolution = 256 x 256
    All inputs and prompt pairs are resized to 256x256 (Sec. IV-A); quantitative scores are resolution-dependent.
assumptions (4)
  • domain assumption Low-level vision tasks are representable as pixel-to-pixel mappings T_task, and an input-target prompt pair [P_S, P_T] is sufficient to communicate the task to a trained model.
    Stated in Sec. III-A, Eq. (2). If visual prompts do not carry enough information to disambiguate tasks, the framework fails.
  • ad hoc to paper Restormer-based multi-task reference models trained per hand-defined group approximate the upper bound of task-specific specialists.
    Introduced in Sec. IV-C.1 as the GLV-Bench evaluation protocol. Only 1-2 specialist models are trained per group, and some reference models fail, e.g., Group 9 dehazing.
  • domain assumption Third-party tools and pretrained models (OpenCV, AdaAttN, and others) produce valid ground-truth targets for synthetic tasks.
    Used in Sec. III-C to construct many of the 101 tasks; if the generated targets are biased, the trained model inherits the bias.
  • domain assumption X-Restormer is a sufficiently general backbone for pixel-level tasks, and bottleneck prompt cross-attention is a sufficient conditioning mechanism.
    Adopted in Sec. III-B; the ablation compares X-Restormer against a ViT backbone, but does not test alternative conditioning placements or backbones.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploring Scalable Unified Modeling for General Low-Level Vision." pith.science (2026). https://pith.science/paper/LC2G4MMV

@misc{pith2026250714801,
  author       = {Pith},
  title        = {Pith review of: Exploring Scalable Unified Modeling for General Low-Level Vision},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LC2G4MMV}},
  note         = {Machine review of arXiv:2507.14801}
}
read the original abstract

Low-level vision involves a wide spectrum of tasks, including image restoration, enhancement, stylization, and feature extraction, which differ significantly in both task formulation and output domains. To address the challenge of unified modeling across such diverse tasks, we propose a Visual task Prompt-based Image Processing (VPIP) framework that leverages input-target image pairs as visual prompts to guide the model in performing a variety of low-level vision tasks. The framework comprises an end-to-end image processing backbone, a prompt encoder, and a prompt interaction module, enabling flexible integration with various architectures and effective utilization of task-specific visual representations. Based on this design, we develop a unified low-level vision model, GenLV, and evaluate its performance across multiple representative tasks. To explore the scalability of this approach, we extend the framework along two dimensions: model capacity and task diversity. We construct a large-scale benchmark consisting of over 100 low-level vision tasks and train multiple versions of the model with varying scales. Experimental results show that the proposed method achieves considerable performance across a wide range of tasks. Notably, increasing the number of training tasks enhances generalization, particularly for tasks with limited data, indicating the model's ability to learn transferable representations through joint training. Further evaluations in zero-shot generalization, few-shot transfer, and task-specific fine-tuning scenarios demonstrate the model's strong adaptability, confirming the effectiveness, scalability, and potential of the proposed framework as a unified foundation for general low-level vision modeling.

Figures

Figures reproduced from arXiv: 2507.14801 by the authors.

Figure 1
Figure 1. Diverse low-level vision tasks. Different categories of [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overall approach of our low-level vision generalist model, GenLV. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Comparison of two attention mechanisms. and z P ΩT through the PCA module. This process yields a task-adapted representation zout, which is subsequently passed through the decoder of the main network to produce the final output image Iout. This design enables the model to condition its behavior on the structure and content of the task prompt, allowing it to implement various tasks. Unlike prior works such as Painter… view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Visual results of different models on various low-level vision tasks. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 6
Figure 6. Figure 6: Results of the mismatch test [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Training loss curves of GenLV variants. on task characteristics (e.g., various denoising types, tone adjustments, or weather-related degradations). For each group, we train a multi-task reference model using the Restormer [35] architecture, which serves as a proxy for …
Figure 8
Figure 8. Figure 8: Examples of GenLV conducting various low-level vision tasks across multiple image domains, compared with the [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Performance of GenLV variants of different scales. [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Performance of GenLV-Huge, reference models, and specialist models. Vertical dashed lines separate task groups. [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Examples of GenLV’s zero-shot performance. Quantitative metrics are shown under the GenLV outputs. The first five [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 12
Figure 12. Figure 12: Performance comparison of different fine-tuning strategies: (b) Tuning the prompt encoder; (c) Tuning the prompt [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]
Figure 13
Figure 13. Figure 13: Examples of GenLV’s few-shot performance. Quan [PITH_FULL_IMAGE:figures/full_fig_p016_13.png]
Figure 15
Figure 15. Figure 15: Failure cases of GenLV. tasks such as image restoration, enhancement, stylization, and feature extraction. By introducing input-target image pairs as visual prompts, the model enables efficient multi-task learning. We construct a large-scale benchmark comprising over …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

67 extracted references · 64 canonical work pages

  1. [1]

    Ai in photography: Scrutinizing implementation of super- resolution techniques in photo-editors,

    N. Fatima, “Ai in photography: Scrutinizing implementation of super- resolution techniques in photo-editors,” in Proc. Int. Conf. Image Vis. Comput. New Zealand , 2020, pp. 1–6. 1

    work page 2020

  2. [2]

    Cardiac image super-resolution with global correspondence using multi-atlas patchmatch,

    W. Shi, J. Caballero, C. Ledig, X. Zhuang, W. Bai, K. Bhatia, A. M. S. M. de Marvao, T. Dawes, D. O’Regan, and D. Rueckert, “Cardiac image super-resolution with global correspondence using multi-atlas patchmatch,” in Proc. Med. Image Comput. Comput.-Assist. Intervent. (MICCAI), 2013, pp. 9–16. 1

    work page 2013

  3. [3]

    Sar image despeckling through convolutional neural networks,

    G. Chierchia, D. Cozzolino, G. Poggi, and L. Verdoliva, “Sar image despeckling through convolutional neural networks,” in Proc. IEEE Int. Geosci. Remote Sens. Symp. (IGARSS) , 2017, pp. 5438–5441. 1

    work page 2017

  4. [4]

    Visual prompt tuning,

    M. Jia, L. Tang, B.-C. Chen, C. Cardie, S. Belongie, B. Hariharan, and S.-N. Lim, “Visual prompt tuning,” in Proc. Eur . Conf. Comput. Vis. (ECCV). Springer, 2022, pp. 709–727. 1, 2

    work page 2022

  5. [5]

    Sequential modeling enables scalable learning for large vision models,

    Y . Bai, X. Geng, K. Mangalam, A. Bar, A. L. Yuille, T. Darrell, J. Malik, and A. A. Efros, “Sequential modeling enables scalable learning for large vision models,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024, pp. 22 861–22 872. 1

    work page 2024

  6. [6]

    All-in-one image restoration for unknown corruption,

    B. Li, X. Liu, P. Hu, Z. Wu, J. Lv, and X. Peng, “All-in-one image restoration for unknown corruption,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2022, pp. 17 452–17 462. 1, 3

    work page 2022

  7. [7]

    Promptir: Prompting for all-in-one image restoration,

    V . Potlapalli, S. W. Zamir, S. Khan, and F. Khan, “Promptir: Prompting for all-in-one image restoration,” in Adv. Neural Inf. Process. Syst. (NeurIPS), 2023. 1, 3

    work page 2023

  8. [8]

    Visual prompting via image inpainting,

    A. Bar, Y . Gandelsman, T. Darrell, A. Globerson, and A. Efros, “Visual prompting via image inpainting,” Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 35, pp. 25 005–25 017, 2022. 1, 2

    work page 2022

Show all 67 references
  1. [9]

    Images speak in images: A generalist painter for in-context visual learning,

    X. Wang, W. Wang, Y . Cao, C. Shen, and T. Huang, “Images speak in images: A generalist painter for in-context visual learning,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2023, pp. 6830–6839. 1, 2, 4

    2023

  2. [10]

    Unifying image processing as visual prompting question answering,

    Y . Liu, X. Chen, X. Ma, X. Wang, J. Zhou, Y . Qiao, and C. Dong, “Unifying image processing as visual prompting question answering,” in Proc. Int. Conf. Mach. Learn. (ICML) . PMLR, 2024, pp. 30 873– 30 891. 1, 2, 4, 5

    2024

  3. [11]

    Learning a low-level vision generalist via visual task prompt,

    X. Chen, Y . Liu, Y . Pu, W. Zhang, J. Zhou, Y . Qiao, and C. Dong, “Learning a low-level vision generalist via visual task prompt,” in Proc. ACM Int. Conf. Multimedia (ACMMM) , 2024, pp. 2671–2680. 2

    2024

  4. [12]

    Learning a deep convolutional network for image super-resolution,

    C. Dong, C. C. Loy, K. He, and X. Tang, “Learning a deep convolutional network for image super-resolution,” in Proc. Eur . Conf. Comput. Vis. (ECCV). Springer, 2014, pp. 184–199. 2

    2014

  5. [13]

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

    K. Zhang, W. Zuo, Y . Chen, D. Meng, and L. Zhang, “Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising,” IEEE Trans. Image Process. , vol. 26, no. 7, pp. 3142–3155, 2017. 2

    2017

  6. [14]

    Defocus deblurring using dual-pixel data,

    A. Abuolaim and M. S. Brown, “Defocus deblurring using dual-pixel data,” in Proc. Eur . Conf. Comput. Vis. (ECCV) . Springer, 2020, pp. 111–126. 2

    2020

  7. [15]

    Compression artifacts reduction by a deep convolutional network,

    C. Dong, Y . Deng, C. C. Loy, and X. Tang, “Compression artifacts reduction by a deep convolutional network,” in Proc. IEEE Int. Conf. Comput. Vis., 2015, pp. 576–584. 2

    2015

  8. [16]

    Deep joint rain detection and removal from a single image,

    W. Yang, R. T. Tan, J. Feng, J. Liu, Z. Guo, and S. Yan, “Deep joint rain detection and removal from a single image,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2017, pp. 1357–1366. 2

    2017

  9. [17]

    Deep joint rain detection and removal from a single image,

    ——, “Deep joint rain detection and removal from a single image,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2017, pp. 1357–1366. 2

    2017

  10. [18]

    A comprehensive overview of image enhancement techniques,

    Y . Qi, Z. Yang, W. Sun, M. Lou, J. Lian, W. Zhao, X. Deng, and Y . Ma, “A comprehensive overview of image enhancement techniques,” Arch. Comput. Methods Eng. , pp. 1–25, 2021. 2

    2021

  11. [19]

    Deep bilateral learning for real-time image enhancement,

    M. Gharbi, J. Chen, J. T. Barron, S. W. Hasinoff, and F. Durand, “Deep bilateral learning for real-time image enhancement,” (TOG), vol. 36, no. 4, pp. 1–12, 2017. 2

    2017

  12. [20]

    Fast local laplacian filters: Theory and applications,

    M. Aubry, S. Paris, S. W. Hasinoff, J. Kautz, and F. Durand, “Fast local laplacian filters: Theory and applications,” (TOG), vol. 33, no. 5, pp. 1–14, 2014. 2, 5

    2014

  13. [21]

    Hdrunet: Single image hdr reconstruction with denoising and dequantization,

    X. Chen, Y . Liu, Z. Zhang, Y . Qiao, and C. Dong, “Hdrunet: Single image hdr reconstruction with denoising and dequantization,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) Workshops , 2021, pp. 354–363. 2

    2021

  14. [22]

    Learning to see in the dark,

    C. Chen, Q. Chen, J. Xu, and V . Koltun, “Learning to see in the dark,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2018, pp. 3291–3300. 2, 13, 14

    2018

  15. [23]

    A computational approach to edge detection,

    J. Canny, “A computational approach to edge detection,” IEEE Trans. Pattern Anal. Mach. Intell. , no. 6, pp. 679–698, 1986. 2

    1986

  16. [24]

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

    J. Johnson, A. Alahi, and L. Fei-Fei, “Perceptual losses for real-time style transfer and super-resolution,” in Proc. Eur . Conf. Comput. Vis. (ECCV). Springer, 2016, pp. 694–711. 2

    2016

  17. [25]

    Language models are few-shot learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askellet al., “Language models are few-shot learners,” Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 33, pp. 1877–1901, 2020. 2

    1901

  18. [26]

    The power of scale for parameter-efficient prompt tuning,

    B. Lester, R. Al-Rfou, and N. Constant, “The power of scale for parameter-efficient prompt tuning,” in Proc. Conf. Empirical Methods Nat. Lang. Process. (EMNLP) . Assoc. Comput. Linguistics, 2021. 2

    2021

  19. [27]

    Lora: Low-rank adaptation of large language models,

    E. J. Hu, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chen et al. , “Lora: Low-rank adaptation of large language models,” in Proc. Int. Conf. Learn. Represent. (ICLR) , 2022. 2

    2022

  20. [28]

    Learning to prompt for vision- language models,

    K. Zhou, J. Yang, C. C. Loy, and Z. Liu, “Learning to prompt for vision- language models,” Int. J. Comput. Vis. , vol. 130, no. 9, pp. 2337–2348,

  21. [29]

    A preliminary exploration towards general image restoration,

    X. Kong, J. Gu, Y . Liu, W. Zhang, X. Chen, Y . Qiao, and C. Dong, “A preliminary exploration towards general image restoration,” arXiv preprint arXiv:2408.15143, 2024. 3

    2024 arXiv

  22. [30]

    Designing a practical degradation model for deep blind image super-resolution,

    K. Zhang, J. Liang, L. Van Gool, and R. Timofte, “Designing a practical degradation model for deep blind image super-resolution,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. , 2021, pp. 4791–4800. 3

    2021

  23. [31]

    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 Proc. IEEE/CVF Int. Conf. Comput. Vis. , 2021, pp. 1905–1914. 3, 6, 13, 15

    2021

  24. [32]

    Unsu- pervised degradation representation learning for blind super-resolution,

    L. Wang, Y . Wang, X. Dong, Q. Xu, J. Yang, W. An, and Y . Guo, “Unsu- pervised degradation representation learning for blind super-resolution,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2021, pp. 10 581–10 590. 3

    2021

  25. [33]

    Prores: Exploring degradation-aware visual prompt for universal image restoration,

    J. Ma, T. Cheng, G. Wang, Q. Zhang, X. Wang, and L. Zhang, “Prores: Exploring degradation-aware visual prompt for universal image restoration,” arXiv preprint arXiv:2306.13653 , 2023. 3

    2023 arXiv

  26. [34]

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

    X. Chen, Z. Li, Y . Pu, Y . Liu, J. Zhou, Y . Qiao, and C. Dong, “A comparative study of image restoration networks for general backbone network design,” in Proc. Eur . Conf. Comput. Vis. (ECCV) . Springer, 2024, pp. 74–91. 4, 6

    2024

  27. [35]

    Restormer: Efficient transformer for high-resolution image restoration,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, and M.-H. Yang, “Restormer: Efficient transformer for high-resolution image restoration,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2022, pp. 5728–5739. 4, 9

    2022

  28. [36]

    Activating more pixels in image super-resolution transformer,

    X. Chen, X. Wang, J. Zhou, Y . Qiao, and C. Dong, “Activating more pixels in image super-resolution transformer,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2023, pp. 22 367–22 377. 4

    2023

  29. [37]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2022, pp. 10 684–10 695. 4

    2022

  30. [38]

    Bayesian-based iterative method of image restora- tion,

    W. H. Richardson, “Bayesian-based iterative method of image restora- tion,” J. Opt. Soc. Am. , vol. 62, no. 1, pp. 55–59, 1972. 5

    1972

  31. [39]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) . IEEE, 2009, pp. 248–255. 5, 13

    2009

  32. [40]

    Benchmarking single-image dehazing and beyond,

    B. Li, W. Ren, D. Fu, D. Tao, D. Feng, W. Zeng, and Z. Wang, “Benchmarking single-image dehazing and beyond,” IEEE Trans. Image Process., vol. 28, no. 1, pp. 492–505, 2018. 5

    2018

  33. [41]

    Multi-scale progressive fusion network for single image deraining,

    K. Jiang, Z. Wang, P. Yi, C. Chen, B. Huang, Y . Luo, J. Ma, and J. Jiang, “Multi-scale progressive fusion network for single image deraining,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2020, pp. 8346–8355. 5

    2020

  34. [42]

    Edge-preserving decompositions for multi-scale tone and detail manipulation,

    Z. Farbman, R. Fattal, D. Lischinski, and R. Szeliski, “Edge-preserving decompositions for multi-scale tone and detail manipulation,” (TOG), vol. 27, no. 3, pp. 1–10, 2008. 5

    2008

  35. [43]

    A new journey from sdrtv to hdrtv,

    X. Chen, Z. Zhang, J. S. Ren, L. Tian, Y . Qiao, and C. Dong, “A new journey from sdrtv to hdrtv,” in Proc. IEEE/CVF Int. Conf. Comput. Vis., 2021, pp. 4500–4509. 5

    2021

  36. [44]

    Deep retinex decomposition for low-light enhancement,

    C. Wei, W. Wang, W. Yang, and J. Liu, “Deep retinex decomposition for low-light enhancement,” in Proc. Brit. Mach. Vis. Conf. (BMVC) ,

  37. [45]

    Learning photo- graphic global tonal adjustment with a database of input/output image pairs,

    V . Bychkovsky, S. Paris, E. Chan, and F. Durand, “Learning photo- graphic global tonal adjustment with a database of input/output image pairs,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) . IEEE, 2011, pp. 97–104. 5

    2011

  38. [46]

    An underwater image enhancement benchmark dataset and beyond,

    C. Li, C. Guo, W. Ren, R. Cong, J. Hou, S. Kwong, and D. Tao, “An underwater image enhancement benchmark dataset and beyond,” IEEE Trans. Image Process. , vol. 29, pp. 4376–4389, 2019. 5 JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 18

    2019

  39. [47]

    Dense extreme inception network: Towards a robust cnn model for edge detection,

    X. S. Poma, E. Riba, and A. Sappa, “Dense extreme inception network: Towards a robust cnn model for edge detection,” in Proc. IEEE/CVF Winter Conf. Appl. Comput. Vis. (WACV) , 2020, pp. 1923–1932. 5

    2020

  40. [48]

    The opencv library,

    G. Bradski, “The opencv library,” Dr . Dobb’s J. Softw. Tools, 2000. 5

    2000

  41. [49]

    Combining sketch and tone for pencil drawing production,

    C. Lu, L. Xu, and J. Jia, “Combining sketch and tone for pencil drawing production,” in Proc. Symp. Non-Photorealistic Animation and Rendering, 2012, pp. 65–73. 5

    2012

  42. [50]

    Structure extraction from texture via relative total variation,

    L. Xu, Q. Yan, Y . Xia, and J. Jia, “Structure extraction from texture via relative total variation,” (TOG), vol. 31, no. 6, pp. 1–10, 2012. 5

    2012

  43. [51]

    Adaattn: Revisit attention mechanism in arbitrary neural style transfer,

    S. Liu, T. Lin, D. He, F. Li, M. Wang, X. Li, Z. Sun, Q. Li, and E. Ding, “Adaattn: Revisit attention mechanism in arbitrary neural style transfer,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. , 2021, pp. 6649–6658. 5

    2021

  44. [52]

    Sars-cov- 2 ct-scan dataset: A large dataset of real patients ct scans for sars-cov-2 identification,

    E. Soares, P. Angelov, S. Biaso, M. H. Froes, and D. K. Abe, “Sars-cov- 2 ct-scan dataset: A large dataset of real patients ct scans for sars-cov-2 identification,” MedRxiv, pp. 2020–04, 2020. 5

    2020

  45. [53]

    Aid: A benchmark data set for performance evaluation of aerial scene classification,

    G.-S. Xia, J. Hu, F. Hu, B. Shi, X. Bai, Y . Zhong, L. Zhang, and X. Lu, “Aid: A benchmark data set for performance evaluation of aerial scene classification,” IEEE Trans. Geosci. Remote Sens. , vol. 55, no. 7, pp. 3965–3981, 2017. 5

    2017

  46. [54]

    Superbench: A super-resolution benchmark dataset for scientific machine learning,

    P. Ren, N. B. Erichson, J. Guo, S. Subramanian, O. San, Z. Lukic, and M. W. Mahoney, “Superbench: A super-resolution benchmark dataset for scientific machine learning,” J. Data-Centric Mach. Learn. Res. , 2025. 5

    2025

  47. [55]

    Rellisur: A real low- light image super-resolution dataset,

    A. Aakerberg, K. Nasrollahi, and T. B. Moeslund, “Rellisur: A real low- light image super-resolution dataset,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS) , 2021. 5

    2021

  48. [56]

    Seal: A framework for systematic evaluation of real-world super- resolution,

    W. Zhang, X. Li, X. Chen, X. Zhang, Y . Qiao, X.-M. Wu, and C. Dong, “Seal: A framework for systematic evaluation of real-world super- resolution,” in Proc. Int. Conf. Learn. Represent. (ICLR) , 2023. 9

    2023

  49. [57]

    Toward convolutional blind denoising of real photographs,

    S. Guo, Z. Yan, K. Zhang, W. Zuo, and L. Zhang, “Toward convolutional blind denoising of real photographs,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2019, pp. 1712–1722. 13, 15

    2019

  50. [58]

    Real-world blur dataset for learning and benchmarking deblurring algorithms,

    J. Rim, H. Lee, J. Won, and S. Cho, “Real-world blur dataset for learning and benchmarking deblurring algorithms,” in Proc. Eur . Conf. Comput. Vis. (ECCV), 2020, pp. 184–201. 13

    2020

  51. [59]

    Multi-stage progressive image restoration,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, M.-H. Yang, and L. Shao, “Multi-stage progressive image restoration,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2021, pp. 14 816–14 826. 13, 15

    2021

  52. [60]

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

    E. Agustsson and R. Timofte, “Ntire 2017 challenge on single image super-resolution: Dataset and study,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) Workshops , 2017, pp. 1122–1131. 13

    2017

  53. [61]

    Aid: A benchmark data set for performance evaluation of aerial scene classification,

    G.-S. Xia, J. Hu, F. Hu, B. Shi, X. Bai, Y . Zhong, L. Zhang, and X. Lu, “Aid: A benchmark data set for performance evaluation of aerial scene classification,” IEEE Trans. Geosci. Remote Sens. , vol. 55, no. 7, pp. 3965–3981, 2017. 13

    2017

  54. [62]

    Continuous remote sensing image super-resolution based on context interaction in implicit function space,

    K. Chen, W. Li, S. Lei, J. Chen, X. Jiang, Z. Zou, and Z. Shi, “Continuous remote sensing image super-resolution based on context interaction in implicit function space,” IEEE Trans. Geosci. Remote Sens., vol. 61, pp. 1–16, 2023. 13, 15

    2023

  55. [63]

    Dis- covering distinctive

    Y . Liu, A. Liu, J. Gu, Z. Zhang, W. Wu, Y . Qiao, and C. Dong, “Dis- covering distinctive ”semantics” in super-resolution networks,” arXiv preprint arXiv:2108.00406, 2021. 13

    2021 arXiv

  56. [64]

    Revisiting the generalization problem of low-level vision models through the lens of image deraining,

    J. Hu, Z. You, J. Gu, K. Zhu, T. Xue, and C. Dong, “Revisiting the generalization problem of low-level vision models through the lens of image deraining,” arXiv preprint arXiv:2502.12600 , 2025. 13

    2025

  57. [65]

    Restoreagent: Autonomous image restoration agent via multimodal large language models,

    H. Chen, W. Li, J. Gu, J. Ren, S. Chen, T. Ye, R. Pei, K. Zhou, F. Song, and L. Zhu, “Restoreagent: Autonomous image restoration agent via multimodal large language models,” in Adv. Neural Inf. Process. Syst. (NeurIPS), 2024. 13

    2024

  58. [66]

    Promptfix: You prompt and we fix the photo,

    Y . Yu, Z. Zeng, H. Hua, J. Fu, and J. Luo, “Promptfix: You prompt and we fix the photo,” in Adv. Neural Inf. Process. Syst. (NeurIPS) , 2024. 16

    2024

  59. [67]

    Pixwizard: Versatile image-to-image visual assistant with open-language instructions,

    W. Lin, X. Wei, R. Zhang, L. Zhuo, S. Zhao, S. Huang, J. Xie, P. Gao, and H. Li, “Pixwizard: Versatile image-to-image visual assistant with open-language instructions,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2025. 16

    2025

Pith tools

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