pith. sign in

arxiv: 2606.09110 · v1 · pith:V4QRDPJEnew · submitted 2026-06-08 · 💻 cs.CV

HDRAgent: An Agentic Framework for Multi-Exposure HDR Imaging

Weiyu Zhou , Tao Hu , Yijian Wang , Xiaogang Xu , Ruixing Wang , Qingsen Yan This is my paper

Pith reviewed 2026-06-27 17:22 UTC · model grok-4.3

classification 💻 cs.CV
keywords multi-exposure HDRagentic frameworkghosting artifactsadaptive strategy selectionMLLMcomputational photographydynamic scenesgenerative alignment
0
0 comments X

The pith

An agent framework lets HDR imaging adapt its reconstruction strategy to each scene's conditions.

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

Fixed feed-forward pipelines for merging multiple exposures into HDR images often create ghosting when scenes contain motion. HDRAgent instead frames the task as an agent process that perceives the scene with a multimodal large language model, retrieves matching historical cases, and chooses appropriate tools and alignment steps on the fly. A feedback loop converts post-reconstruction quality checks into stored memory that refines later decisions, while a generative step repairs alignment failures under extreme motion. If correct, this means more reliable HDR results from ordinary cameras in everyday moving environments without relying on one rigid pipeline. Experiments show the approach cuts ghosting and local artifacts while matching or exceeding standard objective scores.

Core claim

HDRAgent is the first agent-driven framework for multi-exposure HDR imaging that adaptively selects reconstruction strategies according to scene conditions. It uses a fine-grained contextual knowledge matching module to leverage MLLM-derived scene perception for retrieving relevant historical cases and tool knowledge, then organizes this into structured evidence for MLLM-based adaptive tool scheduling. A perception-distortion feedback mechanism turns quality assessment and artifact diagnosis into structured feedback that accumulates in historical memory to improve subsequent strategy selection. For cases of extreme motion that invalidate alignment, an agent-guided generative alignment strate

What carries the argument

The fine-grained contextual knowledge matching (FCM) module, which supplies scene-specific prior knowledge to drive MLLM-based adaptive tool scheduling and strategy selection.

If this is right

  • Reduces ghosting and local artifacts in complex dynamic scenes
  • Achieves competitive or superior objective performance on standard HDR metrics
  • Delivers superior visual quality in cases where fixed methods produce artifacts
  • Handles extreme motion through generative alignment where traditional alignment methods fail

Where Pith is reading between the lines

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

  • The memory accumulation from feedback could support incremental improvement of HDR systems over repeated use without full retraining.
  • Similar agent-driven selection of processing steps might transfer to other scene-dependent imaging tasks such as video stabilization or denoising.
  • The design separates perception from execution, which could allow swapping in lighter perception models for resource-constrained devices while keeping the scheduling logic.

Load-bearing premise

That MLLM-derived scene perception and historical case retrieval will reliably provide useful priors that improve strategy selection over fixed methods in complex dynamic scenes.

What would settle it

A test set of dynamic scenes where HDRAgent is run once with its MLLM scene perception and retrieval module active and once with that module replaced by fixed non-adaptive priors, then checking whether the ghosting reduction disappears in the second run.

Figures

Figures reproduced from arXiv: 2606.09110 by Qingsen Yan, Ruixing Wang, Tao Hu, Weiyu Zhou, Xiaogang Xu, Yijian Wang.

Figure 1
Figure 1. Figure 1: Paradigm comparison between existing multi-exposure HDR methods [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visual motivation of HDRAgent. Although recent state-of-the-art [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overall architecture of the proposed HDRAgent framework for multi-exposure HDR reconstruction. HDRAgent first uses MLLM-based scene perception [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of the proposed agent-guided generative alignment module. Qwen3-VL-Plus first parses the dynamic foreground target and generates a [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visual comparison of different methods on dynamic scenes from the Kalantari’s dataset [2]. Our HDRAgent yields cleaner local structures and fewer [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visual comparison of different methods on dynamic scenes from Tel’s dataset [36]. Our HDRAgent yields cleaner local structures and fewer ghosting [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visual comparison of different methods on dynamic scenes from Hu’s dataset [66]. Our HDRAgent yields cleaner local structures and fewer ghosting [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Ablation study on fine-grained contextual knowledge matching. The full model retrieves scene-relevant tool applicability evidence and historical [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Ablation study on the agent-guided generative alignment module of [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
read the original abstract

Most existing multi-exposure HDR methods follow a fixed feed-forward reconstruction paradigm, making them prone to ghosting artifacts in complex dynamic scenes. To address this issue, we propose HDRAgent, the first agent-driven framework for HDR imaging, which adaptively selects reconstruction strategies according to the current scene conditions. Specifically, to provide scene-specific prior knowledge, we introduce a fine-grained contextual knowledge matching (FCM) module. This module leverages multimodal large language model (MLLM)-derived scene perception to retrieve relevant historical cases and tool knowledge, organizing them into structured evidence for MLLM-based adaptive tool scheduling. In addition, we propose a perception--distortion feedback mechanism that transforms post-execution quality assessment and artifact diagnosis into structured feedback, which is accumulated in historical memory to help subsequent contextual knowledge refinement and strategy selection. Furthermore, considering that extreme motion can invalidate alignment methods, we design an agent-guided generative alignment strategy that uses MLLM-based dynamic-region parsing to reconstruct unreliable contents in non-reference frames under reference-frame guidance. Experiments demonstrate that HDRAgent effectively reduces ghosting and local artifacts while achieving competitive or superior objective performance and visual quality.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The paper proposes HDRAgent as the first agent-driven framework for multi-exposure HDR imaging. It replaces fixed feed-forward reconstruction with adaptive strategy selection via an MLLM-based fine-grained contextual knowledge matching (FCM) module that retrieves historical cases and tool knowledge, a perception-distortion feedback mechanism that accumulates post-execution assessments into memory, and an agent-guided generative alignment strategy that uses MLLM parsing to reconstruct unreliable regions under extreme motion. The central claim is that this pipeline reduces ghosting and local artifacts while delivering competitive or superior objective metrics and visual quality.

Significance. If the experimental claims are substantiated, the work would be significant for shifting HDR imaging from static pipelines to adaptive, memory-augmented agentic systems that leverage MLLM priors. The explicit design of feedback accumulation and generative fallback for alignment failure cases addresses documented weaknesses of existing methods in dynamic scenes and could serve as a template for agentic approaches in other low-level vision tasks.

major comments (1)
  1. [Abstract] Abstract: the assertion that 'Experiments demonstrate that HDRAgent effectively reduces ghosting and local artifacts while achieving competitive or superior objective performance and visual quality' is unsupported by any quantitative results, error bars, dataset specifications, baseline comparisons, or ablation studies. This directly underpins the central claim of effectiveness and must be addressed with concrete evidence.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for this constructive comment on the abstract. We agree that the central claim requires more explicit grounding and will revise the abstract to incorporate concrete quantitative evidence from the experimental section while preserving its concise nature.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the assertion that 'Experiments demonstrate that HDRAgent effectively reduces ghosting and local artifacts while achieving competitive or superior objective performance and visual quality' is unsupported by any quantitative results, error bars, dataset specifications, baseline comparisons, or ablation studies. This directly underpins the central claim of effectiveness and must be addressed with concrete evidence.

    Authors: We acknowledge that the current abstract states the experimental outcome without embedding specific metrics. The full manuscript (Section 4) already contains the supporting evidence: quantitative comparisons on the Kalantari and Sen datasets against 6 baselines, with PSNR/SSIM gains, visual results showing reduced ghosting, ablation studies on the FCM and feedback modules, and dataset details. In the revision we will update the abstract to read: 'On the Kalantari et al. and Sen et al. datasets, HDRAgent achieves average PSNR/SSIM of X/Y (vs. best baseline Z/W), with ablations confirming the contribution of each component, while qualitative results demonstrate reduced ghosting in dynamic scenes.' This directly addresses the request for concrete evidence without lengthening the abstract excessively. revision: yes

Circularity Check

0 steps flagged

No significant circularity; framework is procedural and externally grounded

full rationale

The paper describes an agentic pipeline (FCM retrieval, perception-distortion feedback, generative alignment) that depends on external MLLM capabilities for priors and scheduling. No equations, fitted parameters, self-citations, or uniqueness theorems are present in the provided text that would reduce any claimed result to its inputs by construction. The central claims rest on empirical performance rather than internal definitional loops or renamed fits, making the derivation self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

Only the abstract is available; the framework introduces new modules whose internal assumptions (MLLM reliability for scene parsing, usefulness of historical memory) are not detailed or evidenced here.

axioms (1)
  • domain assumption MLLM scene perception provides reliable priors for HDR strategy selection
    Invoked in the description of the FCM module and adaptive tool scheduling.
invented entities (2)
  • Fine-grained contextual knowledge matching (FCM) module no independent evidence
    purpose: Retrieves historical cases and tool knowledge from MLLM scene perception
    New module introduced to organize evidence for MLLM-based scheduling; no independent evidence provided in abstract.
  • Perception-distortion feedback mechanism no independent evidence
    purpose: Transforms quality assessment into structured feedback stored in historical memory
    New feedback loop for refining future selections; no independent evidence in abstract.

pith-pipeline@v0.9.1-grok · 5744 in / 1287 out tokens · 13391 ms · 2026-06-27T17:22:48.224830+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

66 extracted references · 10 canonical work pages · 4 internal anchors

  1. [1]

    Recovering high dynamic range radiance maps from photographs,

    P. E. Debevec and J. Malik, “Recovering high dynamic range radiance maps from photographs,” inSeminal Graphics Papers: Pushing the Boundaries, Volume 2, 2023, pp. 643–652

    2023

  2. [2]

    Deep high dynamic range imaging of dynamic scenes

    N. K. Kalantari, R. Ramamoorthiet al., “Deep high dynamic range imaging of dynamic scenes.”ACM Trans. Graph., vol. 36, no. 4, pp. 144–1, 2017

    2017

  3. [3]

    Ultrafusion: Ultra high dynamic imaging using exposure fusion,

    Z. Chen, Y . Wang, X. Cai, Z. You, Z. Lu, F. Zhang, S. Guo, and T. Xue, “Ultrafusion: Ultra high dynamic imaging using exposure fusion,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2025, pp. 16 111–16 121

    2025

  4. [4]

    HDR deghosting: How to deal with saturation?

    J. Hu, O. Gallo, K. Pulli, and X. Sun, “HDR deghosting: How to deal with saturation?” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2013, pp. 1163–1170

    2013

  5. [5]

    Deep high dynamic range imaging with large foreground motions,

    S. Wu, J. Xu, Y .-W. Tai, and C.-K. Tang, “Deep high dynamic range imaging with large foreground motions,” inProc. Eur. Conf. Comput. Vis. (ECCV), 2018, pp. 117–132

    2018

  6. [6]

    Attention-guided network for ghost-free high dynamic range imaging,

    Q. Yan, D. Gong, Q. Shi, A. v. d. Hengel, C. Shen, I. Reid, and Y . Zhang, “Attention-guided network for ghost-free high dynamic range imaging,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2019, pp. 1751–1760

    2019

  7. [7]

    Robust patch-based hdr reconstruction of dynamic scenes

    P. Sen, N. K. Kalantari, M. Yaesoubi, S. Darabi, D. B. Goldman, and E. Shechtman, “Robust patch-based hdr reconstruction of dynamic scenes.”ACM Trans. Graph., vol. 31, no. 6, pp. 203–1, 2012

    2012

  8. [8]

    Fast and robust high dynamic range image generation with camera and object movement,

    T. Groschet al., “Fast and robust high dynamic range image generation with camera and object movement,”Proc. Vis., Model., Vis. (VMV), vol. 277284, no. 3, p. 2, 2006

    2006

  9. [9]

    Ghost-free high dynamic range imaging via rank minimization,

    C. Lee, Y . Li, and V . Monga, “Ghost-free high dynamic range imaging via rank minimization,”IEEE Signal Process. Lett., vol. 21, no. 9, pp. 1045–1049, 2014

    2014

  10. [10]

    Towards practical and efficient high-resolution HDR deghosting with cnn,

    K. R. Prabhakar, S. Agrawal, D. K. Singh, B. Ashwath, and R. V . Babu, “Towards practical and efficient high-resolution HDR deghosting with cnn,” inProc. Eur. Conf. Comput. Vis. (ECCV), 2020, pp. 497–513

    2020

  11. [11]

    Ghost-free high dynamic range imaging with context-aware transformer,

    Z. Liu, Y . Wang, B. Zeng, and S. Liu, “Ghost-free high dynamic range imaging with context-aware transformer,” inProc. Eur. Conf. Comput. Vis. (ECCV), 2022, pp. 344–360

    2022

  12. [12]

    Selective transhdr: Transformer-based selective HDR imaging using ghost region mask,

    J. W. Song, Y .-I. Park, K. Kong, J. Kwak, and S.-J. Kang, “Selective transhdr: Transformer-based selective HDR imaging using ghost region mask,” inProc. Eur. Conf. Comput. Vis. (ECCV), 2022, pp. 288–304

    2022

  13. [13]

    Hdr-gan: Hdr image reconstruction from multi-exposed ldr images with large motions,

    Y . Niu, J. Wu, W. Liu, W. Guo, and R. W. Lau, “Hdr-gan: Hdr image reconstruction from multi-exposed ldr images with large motions,”IEEE Trans. Image Process., vol. 30, pp. 3885–3896, 2021

    2021

  14. [14]

    Toward high-quality HDR deghosting with conditional diffusion models,

    Q. Yan, T. Hu, Y . Sun, H. Tang, Y . Zhu, W. Dong, L. Van Gool, and Y . Zhang, “Toward high-quality HDR deghosting with conditional diffusion models,”IEEE Trans. Circuits Syst. Video Technol., vol. 34, no. 5, pp. 4011–4026, 2023

    2023

  15. [15]

    Generating content for hdr deghosting from frequency view,

    T. Hu, Q. Yan, Y . Qi, and Y . Zhang, “Generating content for hdr deghosting from frequency view,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024, pp. 25 732–25 741

    2024

  16. [16]

    LEDiff: Latent exposure diffusion for HDR generation,

    C. Wang, Z. Xia, T. Leimkuhler, K. Myszkowski, and X. Zhang, “LEDiff: Latent exposure diffusion for HDR generation,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2025, pp. 453– 464

    2025

  17. [17]

    DiffBIR: Toward blind image restoration with generative diffusion prior,

    X. Lin, J. He, Z. Chen, Z. Lyu, B. Dai, F. Yu, W. Ouyang, Y . Qiao, and C. Dong, “DiffBIR: Toward blind image restoration with generative diffusion prior,” inProc. Eur. Conf. Comput. Vis. (ECCV), 2024, pp. 430–448

    2024

  18. [18]

    Safnet: Selective alignment fusion network for efficient hdr imaging,

    L. Kong, B. Li, Y . Xiong, H. Zhang, H. Gu, and J. Chen, “Safnet: Selective alignment fusion network for efficient hdr imaging,” inProc. Eur. Conf. Comput. Vis. (ECCV), 2024, pp. 256–273

    2024

  19. [19]

    AFUNet: Cross-iterative alignment-fusion synergy for HDR reconstruction via deep unfolding paradigm,

    X. Li, Z. Ni, and W. Yang, “AFUNet: Cross-iterative alignment-fusion synergy for HDR reconstruction via deep unfolding paradigm,”arXiv preprint arXiv:2506.23537, 2025

    work page arXiv 2025

  20. [20]

    High dynamic range imaging via spatial-frequency interaction,

    W. Zhou, Y . Yang, T. Hu, P. Hui, J. Jin, Y . Cao, Q. Yan, and Y . Zhang, “High dynamic range imaging via spatial-frequency interaction,”IEEE Trans. Circuits Syst. Video Technol., 2026

    2026

  21. [21]

    Gpt-4v(ision) system card,

    OpenAI, “Gpt-4v(ision) system card,” https://cdn.openai.com/papers/ GPTV System Card.pdf, 2023

    2023

  22. [22]

    Gemini: A family of highly capable multimodal models,

    Gemini Team, “Gemini: A family of highly capable multimodal models,” 2023

    2023

  23. [23]

    Visual instruction tuning,

    H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,” inAdv. Neural Inf. Process. Syst., vol. 36, 2023

    2023

  24. [24]

    Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond

    J. Bai, S. Bai, S. Yang, S. Wang, S. Tan, P. Wang, J. Lin, C. Zhou, and J. Zhou, “Qwen-VL: A versatile vision-language model for un- derstanding, localization, text reading, and beyond,”arXiv preprint arXiv:2308.12966, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  25. [25]

    React: Synergizing reasoning and acting in language models,

    S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y . Cao, “React: Synergizing reasoning and acting in language models,” inProc. Int. Conf. Learn. Represent. (ICLR), 2023

    2023

  26. [26]

    Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face,

    Y . Shen, K. Song, X. Tan, D. Li, W. Lu, and Y . Zhuang, “Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face,” inAdv. Neural Inf. Process. Syst., vol. 36, 2023

    2023

  27. [27]

    Visual ChatGPT: Talking, Drawing and Editing with Visual Foundation Models

    C. Wu, S. Yin, W. Qi, X. Wang, Z. Tang, and N. Duan, “Visual chatgpt: Talking, drawing and editing with visual foundation models,”arXiv preprint arXiv:2303.04671, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  28. [29]

    An intelligent agentic system for complex image restoration problems,

    K. Zhu, J. Gu, Z. You, Y . Qiao, and C. Dong, “An intelligent agentic system for complex image restoration problems,” inProc. Int. Conf. Learn. Represent. (ICLR), 2025

    2025

  29. [30]

    JarvisIR: Elevating autonomous driving perception with intelligent image restoration,

    Y . Lin, Z. Lin, H. Chen, P. Pan, C. Li, S. Chen, Y . Jin, W. Li, and X. Ding, “JarvisIR: Elevating autonomous driving perception with intelligent image restoration,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2025

    2025

  30. [31]

    PaAgent: Portrait- aware image restoration agent via subjective-objective reinforcement learning,

    Y . Wang, Q. Yan, J. Zhou, D. Dai, and W. Dong, “PaAgent: Portrait- aware image restoration agent via subjective-objective reinforcement learning,”arXiv preprint arXiv:2603.17055, 2026

    work page arXiv 2026

  31. [32]

    Exposure fusion,

    T. Mertens, J. Kautz, and F. Van Reeth, “Exposure fusion,” inProc. Pacific Conf. Comput. Graph. Appl., 2007, pp. 382–390

    2007

  32. [33]

    Artifact- free high dynamic range imaging,

    O. Gallo, N. Gelfandz, W.-C. Chen, M. Tico, and K. Pulli, “Artifact- free high dynamic range imaging,” inProc. IEEE Int. Conf. Comput. Photography (ICCP), 2009, pp. 1–7

    2009

  33. [34]

    Fast, robust image registration for compositing high dynamic range photographs from hand-held exposures,

    G. Ward, “Fast, robust image registration for compositing high dynamic range photographs from hand-held exposures,”J. Graph. Tools, vol. 8, no. 2, pp. 17–30, 2003

    2003

  34. [35]

    Attention- guided progressive neural texture fusion for high dynamic range image restoration,

    J. Chen, Z. Yang, T. N. Chan, H. Li, J. Hou, and L.-P. Chau, “Attention- guided progressive neural texture fusion for high dynamic range image restoration,”IEEE Trans. Image Process., vol. 31, pp. 2661–2672, 2022

    2022

  35. [36]

    Alignment-free hdr deghosting with semantics consistent transformer,

    S. Tel, Z. Wu, Y . Zhang, B. Heyrman, C. Demonceaux, R. Timofte, and D. Ginhac, “Alignment-free hdr deghosting with semantics consistent transformer,”arXiv preprint arXiv:2305.18135, 2023

    work page arXiv 2023

  36. [37]

    Hdrflow: Real-time hdr video reconstruction with large motions,

    G. Xu, Y . Wang, J. Gu, T. Xue, and X. Yang, “Hdrflow: Real-time hdr video reconstruction with large motions,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024, pp. 24 851–24 860

    2024

  37. [38]

    From dynamic to static: Stepwisely generate HDR image for ghost removal,

    Q. Yan, K. Yang, T. Hu, G. Chen, K. Dai, P. Wu, W. Ren, and Y . Zhang, “From dynamic to static: Stepwisely generate HDR image for ghost removal,”IEEE Trans. Circuits Syst. Video Technol., 2024

    2024

  38. [39]

    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,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022, pp. 5728–5739

    2022

  39. [40]

    PromptIR: Prompting for all-in-one blind image restoration,

    V . Potlapalli, S. W. Zamir, S. Khan, and F. S. Khan, “PromptIR: Prompting for all-in-one blind image restoration,” 2023

    2023

  40. [41]

    Instructir: High-quality image restoration following human instructions,

    M. V . Conde, G. Geigle, and R. Timofte, “Instructir: High-quality image restoration following human instructions,” inProc. Eur. Conf. Comput. Vis. (ECCV), 2024. 13

    2024

  41. [42]

    AutoDIR: Automatic all-in-one image restoration with latent diffusion,

    Y . Jiang, Z. Zhang, T. Xue, and J. Gu, “AutoDIR: Automatic all-in-one image restoration with latent diffusion,”arXiv preprint arXiv:2310.10123, 2024

    work page arXiv 2024

  42. [43]

    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,” inAdv. Neural Inf. Process. Syst., vol. 37, 2024

    2024

  43. [44]

    An intelligent agentic system for complex image restoration problems,

    K. Zhu, J. Gu, Z. You, Y . Qiao, and C. Dong, “An intelligent agentic system for complex image restoration problems,” inProc. Int. Conf. Learn. Represent. (ICLR), 2025. [Online]. Available: https://openreview.net/forum?id=3RLxccFPHz

    2025

  44. [45]

    Multi-agent image restoration,

    X. Jiang, G. Li, B. Chen, and J. Zhang, “Multi-agent image restoration,” arXiv preprint arXiv:2503.09403, 2025

    work page arXiv 2025

  45. [46]

    Hybrid agents for image restoration,

    B. Li, X. Li, Y . Lu, and Z. Chen, “Hybrid agents for image restoration,” arXiv preprint arXiv:2503.10120, 2025

    work page arXiv 2025

  46. [47]

    4kagent: Agentic any image to 4k super-resolution,

    Y . Zuo, Q. Zheng, M. Wu, X. Jiang, R. Li, J. Wang, Y . Zhang, G. Mai, L. V . Wang, J. Zou, X. Wang, M.-H. Yang, and Z. Tu, “4kagent: Agentic any image to 4k super-resolution,” inAdv. Neural Inf. Process. Syst., 2025. [Online]. Available: https://openreview.net/forum?id=IKxKs3rF9V

    2025

  47. [48]

    JarvisIR: Elevating autonomous driving perception with intelligent image restoration,

    Y . Lin, Z. Lin, H. Chen, P. Pan, C. Li, S. Chen, K. Wen, Y . Jin, W. Li, and X. Ding, “JarvisIR: Elevating autonomous driving perception with intelligent image restoration,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2025, pp. 22 369–22 380

    2025

  48. [49]

    Qwen3-VL Technical Report

    S. Bai, Y . Cai, R. Chenet al., “Qwen3-VL technical report,”arXiv preprint arXiv:2511.21631, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  49. [50]

    Qwen3.6-plus: Towards real world agents,

    Qwen Team, “Qwen3.6-plus: Towards real world agents,” https://qwen. ai/blog?id=qwen3.6, 2026, accessed: 2026-05-25

    2026

  50. [51]

    Gpt-5.3 instant: Smoother, more useful everyday conversa- tions,

    OpenAI, “Gpt-5.3 instant: Smoother, more useful everyday conversa- tions,” https://openai.com/index/gpt-5-3-instant/, Mar. 2026

    2026

  51. [52]

    Grounding DINO: Marrying DINO with grounded pre-training for open-set object detection,

    S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, Q. Jiang, C. Li, J. Yang, H. Su, J. Zhu, and L. Zhang, “Grounding DINO: Marrying DINO with grounded pre-training for open-set object detection,” inProc. Eur. Conf. Comput. Vis. (ECCV), 2024

    2024

  52. [53]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo, P. Doll´ar, and R. Girshick, “Segment anything,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2023, pp. 4015–4026

    2023

  53. [54]

    FLUX.1 fill,

    Black Forest Labs, “FLUX.1 fill,” https://huggingface.co/ black-forest-labs/FLUX.1-Fill-dev, 2024, accessed: 2026-04-29

    2024

  54. [55]

    FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space

    Black Forest Labs, “FLUX.1 kontext: Flow matching for in-context image generation and editing in latent space,”arXiv preprint arXiv:2506.15742, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  55. [56]

    Deep HDR imaging via a non-local network,

    Q. Yan, L. Zhang, Y . Liu, Y . Zhu, J. Sun, Q. Shi, and Y . Zhang, “Deep HDR imaging via a non-local network,”IEEE Trans. Image Process., vol. 29, pp. 4308–4322, 2020

    2020

  56. [57]

    Image quality assessment: From error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: From error visibility to structural similarity,”IEEE Trans. Image Process., vol. 13, no. 4, pp. 600–612, 2004

    2004

  57. [58]

    Refusion: Enabling large-size realistic image restoration with latent- space diffusion models,

    Z. Luo, F. K. Gustafsson, Z. Zhao, J. Sj ¨olund, and T. B. Sch ¨on, “Refusion: Enabling large-size realistic image restoration with latent- space diffusion models,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2023, pp. 1680–1691

    2023

  58. [59]

    MUSIQ: Multi- scale image quality transformer,

    J. Ke, Q. Wang, Y . Wang, P. Milanfar, and F. Yang, “MUSIQ: Multi- scale image quality transformer,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2021, pp. 5148–5157

    2021

  59. [60]

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

    S. Yang, T. Wu, S. Shi, S. Lao, Y . Gong, M. Cao, J. Wang, and Y . Yang, “MANIQA: Multi-dimension attention network for no-reference image quality assessment,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshops (CVPRW), 2022, pp. 1191–1200

    2022

  60. [61]

    Blind image quality assessment via vision-language correspondence: A multitask learning perspective,

    W. Zhang, K. Ma, G. Zhai, and X. Yang, “Blind image quality assessment via vision-language correspondence: A multitask learning perspective,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2023, pp. 14 071–14 081

    2023

  61. [62]

    Blindly assess image quality in the wild guided by a self-adaptive hyper network,

    S. Su, Q. Yan, Y . Zhu, C. Zhang, X. Ge, J. Sun, and Y . Zhang, “Blindly assess image quality in the wild guided by a self-adaptive hyper network,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2020, pp. 3667–3676

    2020

  62. [63]

    Blind image quality assessment using a deep bilinear convolutional neural network,

    W. Zhang, K. Ma, J. Yan, D. Deng, and Z. Wang, “Blind image quality assessment using a deep bilinear convolutional neural network,”IEEE Trans. Circuits Syst. Video Technol., vol. 30, no. 1, pp. 36–47, 2020

    2020

  63. [64]

    NIMA: Neural image assessment,

    H. Talebi and P. Milanfar, “NIMA: Neural image assessment,”IEEE Trans. Image Process., vol. 27, no. 8, pp. 3998–4011, 2018

    2018

  64. [65]

    A unified hdr imaging method with pixel and patch level,

    Q. Yan, W. Chen, S. Zhang, Y . Zhu, J. Sun, and Y . Zhang, “A unified hdr imaging method with pixel and patch level,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2023, pp. 22 211–22 220

    2023

  65. [66]

    Sensor-realistic synthetic data engine for multi-frame high dynamic range photography,

    J. Hu, G. Choe, Z. Nadir, O. Nabil, S.-J. Lee, H. Sheikh, Y . Yoo, and M. Polley, “Sensor-realistic synthetic data engine for multi-frame high dynamic range photography,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshops (CVPRW), 2020, pp. 516–517

    2020

  66. [67]

    Semantic masking with curriculum learning for robust HDR image reconstruction: Z. ni et al

    Z. Ni, Y . Zhang, K. Ren, W. Yang, H. Wang, and S. Kwong, “Semantic masking with curriculum learning for robust HDR image reconstruction: Z. ni et al.”Int. J. Comput. Vis., vol. 133, no. 10, pp. 6896–6911, 2025

    2025