Pith. sign in

REVIEW 4 major objections 8 minor 3 cited by

A trained vision-language agent learns to call the right restoration tool at each step, beating exhaustive agents and all-in-one models while running over 2.5× faster.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-13 16:46 UTC pith:HIFQ4WZL

load-bearing objection Solid first trainable VLM agent for IR tool orchestration; SOTA + 2.5× speed claims hold on the reported tables, but rest almost entirely on MiO100 and 4KAgent-derived demos. the 4 major comments →

arxiv 2603.27742 v2 pith:HIFQ4WZL submitted 2026-03-29 cs.CV

TIR-Agent: Training an Explorative and Efficient Agent for Image Restoration

Guoli Jia , Yisheng Zhang , Haote Hu , Shanxu Zhao , Kaikai Zhao , Long Sun , Xinwei Long , Kai Tian
show 6 more authors
Che Jiang Zhaoxiang Liu Kai Wang Shiguo Lian Kaiyan Zhang Bowen Zhou
This is my paper
classification cs.CV
keywords image restorationvision-language agentreinforcement learningtool callingcomposite degradationmulti-metric rewardexploration
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Most image-restoration agents still plan by hand or exhaustively try every specialist model for every sub-task. That produces both wasteful computation and brittle ordering of steps such as denoise-then-dehaze. TIR-Agent instead trains a vision-language model to decide, from the current image and history, which restoration task and which specialist tool to invoke next. Training proceeds in two stages: supervised fine-tuning on demonstration trajectories that have been deliberately scrambled in order and tool choice, followed by reinforcement learning whose reward is a dynamically re-weighted mix of full-reference and no-reference quality metrics. The result is a policy that restores composite degradations more accurately than twelve published baselines—including six all-in-one networks, three training-free agents and three large proprietary models—while eliminating the redundant tool executions that dominate inference cost.

Core claim

The central claim is that a vision-language model, once given a base policy broadened by random task-order and tool-choice perturbation and then refined by a multi-dimensional adaptive reward, can learn a direct tool-calling policy that both improves restoration quality on mixed degradations and removes the need for exhaustive tool search, yielding more than 2.5× faster inference than training-free agents.

What carries the argument

Exploration-Driven Perturbation (EDP) plus Multi-dimensional Adaptive Reward (MAR): EDP injects controlled randomness into SFT trajectories so the policy explores a wider space of schedules and tools; MAR re-weights heterogeneous image-quality metrics on the fly so no single metric is gamed during RL.

Load-bearing premise

The claim rests on the premise that demonstration trajectories produced by an existing training-free agent, after random order and tool scrambling, still form a rich enough starting policy for reinforcement learning to reach a better global solution rather than a local trap.

What would settle it

Retrain TIR-Agent from the same SFT data but with the random-order and random-tool perturbations removed, then measure whether final PSNR/SSIM and no-reference scores on the MiO100 out-of-domain splits fall back to or below the unperturbed baseline; a large drop would falsify the necessity of the exploration step.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 8 minor

Summary. The paper proposes TIR-Agent, a vision-language image-restoration agent (Qwen3-VL-8B) trained with SFT then RL to learn a direct task-and-tool calling policy, replacing the heuristic scheduling and exhaustive tool traversal used by training-free IR agents. Two algorithmic components are introduced: Exploration-Driven Perturbation (EDP; Eqs. 1–4), which randomly reorders tasks and mixes tool preferences in 4KAgent-generated demonstration trajectories, and Multi-Dimensional Adaptive Reward (MAR; Eqs. 5–9), which EMA-based reweights and decouples FR/NR metrics to reduce reward hacking. An engineering MC-Pool supports concurrent GPU tool calls during RL. On MiO100 (Groups A–C) and synthetic out-of-domain mixtures (D=2, D≥3), the method is reported to beat 6 all-in-one models, 3 training-free agents, 3 proprietary VLMs, and SimpleCall on most metrics, with ≥2.5× inference speedup versus AgenticIR/4KAgent.

Significance. If the results hold under broader evaluation, this is a solid systems-and-methods contribution: it is among the first systematic applications of agentic RL to composite image restoration with a full VLM policy, and it directly targets the two practical failures of training-free agents (suboptimal ordering and linear tool enumeration). The efficiency claim is well supported by wall-clock times in Tables 1–3 and is of clear practical value. EDP and MAR are concrete, ablated designs (Table 5, Figs. 6–7) rather than pure prompting tricks; the MC-Pool is a useful training infrastructure note. Strengths include multi-group quantitative tables, proprietary-model analysis (Fig. 4), qualitative examples, and ablations that isolate SFT, RL, α_t, α_m, decoupling, and dynamic weights. The work is significant for modular IR systems even if absolute SOTA claims need tighter qualification.

major comments (4)
  1. The central SOTA/generalization claim rests almost entirely on MiO100 (Tables 1–4), a synthetic composite-degradation benchmark from the same agentic-IR line that supplies the SFT demonstrations (4KAgent [58]; §4.2). RL data are ImageNet-derived but degraded “following [57]” with only a depth-estimator swap, and OOD splits (D=2 / D≥3) remain inside that synthesis protocol. This does not make the results circular, but it does make the ranking vulnerable to shared degradation synthesis, tool set, and metric suite. For a claim of outperforming 12 baselines “on both in-domain and out-of-domain degradations,” at least one independent real-world or alternate-synthesis benchmark (e.g., real multi-degradation captures, or a protocol not tied to AgenticIR/4KAgent) is load-bearing; without it the generalization language should be substantially narrowed.
  2. The abstract and §4.3 assert that TIR-Agent “outperforms 12 baselines” including proprietary models. Table 1 Group C shows 4KAgent higher PSNR (19.77 vs 19.53); Table 3 Group C shows Claude Opus 4.5 and Gemini 3.0 Pro higher PSNR/SSIM (e.g., 20.27/0.6043 and 20.38/0.6118 vs 19.53/0.5643). Gains are strongest and most consistent on NR metrics and on Groups A/B. The blanket outperform claim is therefore not supported as stated and should be restated metric- and group-wise (e.g., “best or competitive on NR and most FR settings, with mixed FR results on Group C”).
  3. §3 and Eqs. 1–4 take as given that random order/tool perturbations of 4KAgent trajectories expand support without systematically harmful demos. Fig. 7 and the α_t ablation show that α_t>0.3 sharply hurts PSNR, which the text attributes to “low-quality images appear[ing] before high-quality ones.” That fragility is a load-bearing assumption for the SFT→RL pipeline. The paper needs either (i) a quantitative check that EDP trajectories improve final restored quality under FR metrics before RL, or (ii) an explicit discussion of when order perturbation injects invalid or anti-correlated demonstrations, plus sensitivity beyond the chosen α_t=0.3, α_m=0.4 operating point.
  4. Comparison to the only other trainable agent, SimpleCall (Table 2), is restricted to five Group-A combinations and partial published numbers, with SimpleCall not publicly available. Given that SimpleCall is the closest methodological baseline, the superiority claim over “trainable agents” is under-supported. Either expand the comparison protocol (same tools/degradations, full metric suite) or clearly limit the claim to training-free agents plus all-in-one models.
minor comments (8)
  1. Conclusion: typo “a trainable vision-language agent that that addresses”.
  2. §4.3 baselines: “Claude Pous 4.5” should be “Claude Opus 4.5”.
  3. §4.4 / §4.5 cross-references appear broken: EDP order/tool analysis is discussed as “Fig. 4 (c)/(d)” while Fig. 4 is the proprietary-model analysis; the EDP/α plots are Fig. 7. Please renumber and re-cite consistently.
  4. §4.5: “entorpy” → “entropy”.
  5. Table 2 notation “SimpleCall rr / fr” is unexplained in the caption; define the reward variants.
  6. Eqs. 5–7: the EMA symbol is written identically to the batch reward (r_θ1); use a distinct bar/tilde notation for the EMA state to avoid confusion.
  7. §3.4 MC-Pool is described at a high level; a short note on failure rates, retry impact, or scaling with b×g would help readers reproduce the training setup.
  8. Related Work: SimpleCall [30] is dated 2025/arXiv; ensure citation and capability claims match the public version used for comparison.

Circularity Check

0 steps flagged

No circularity: empirical SFT+RL agent paper; rewards and benchmarks are external, and demonstration trajectories do not force the claimed gains by construction.

full rationale

TIR-Agent is a standard two-stage empirical ML system (SFT on demonstration trajectories, then RL against FR/NR image-quality metrics). The SFT seeds come from the external training-free pipeline 4KAgent [58] (no author overlap with the present paper), then are deliberately perturbed by EDP (Eqs. 1–4) and refined by RL with MAR (Eqs. 5–9) on ImageNet-derived pairs stated to be disjoint from the test set. Final claims are comparative numbers on MiO100 against all-in-one models, other agents, and proprietary VLMs—not first-principles predictions. No quantity is defined in terms of the quantity it is said to predict; no fitted scalar is renamed a prediction; no uniqueness theorem or ansatz is imported from the authors’ prior work to force the result. Concerns that MiO100 shares a degradation-synthesis lineage with the demonstration generator are about evaluation independence / generalization risk, not about a derivation that reduces by construction to its inputs. Score 0 is therefore the correct finding.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 3 invented entities

The central claim rests on standard deep-learning and RL practice plus a handful of paper-specific design choices (perturbation rates, EMA/clipping, metric set) and the assumption that 4KAgent-generated trajectories after random perturbation are a usable base policy. No new physical entities are postulated; the invented items are algorithmic modules.

free parameters (4)
  • α_t (task-order perturbation probability) = 0.3
    Hand-chosen after sensitivity sweep; set to 0.3 because higher values degrade PSNR (Fig. 7a). Directly controls SFT trajectory diversity that RL later exploits.
  • α_m (tool-selection perturbation probability) = 0.4
    Hand-chosen after sensitivity sweep; set to 0.4 to raise tool entropy without harming PSNR (Fig. 7c). Controls exploration of specialist models per task.
  • EMA decay β and clip ε for MAR weights
    Smoothing and clipping constants that determine how aggressively under-performing metrics are up-weighted; values affect the multi-objective balance shown in Fig. 6.
  • SFT/RL learning rates, batch size, rollout group size g = 1e-5 / 5e-6 / 64 / 8
    Standard training hyper-parameters (1e-5 SFT, 5e-6 RL, batch 64, g=8) that shape the final policy; not derived from first principles.
axioms (4)
  • domain assumption Composite image quality can be adequately summarized by the chosen FR (PSNR, SSIM, LPIPS) and NR (MANIQA, CLIPIQA, MUSIQ) metrics for both reward and evaluation.
    Invoked throughout §3.3 and §4; the entire MAR design and all tables rest on this metric suite.
  • domain assumption A pretrained VLM (Qwen3-VL-8B) can acquire degradation perception and tool-calling format from a few tens of thousands of demonstration trajectories.
    Stated in §3 and confirmed by the ‘w/o SFT’ ablation; without it the RL stage has no usable base policy.
  • ad hoc to paper Random order and tool perturbations of 4KAgent trajectories expand the support of the policy without introducing systematically harmful demonstrations.
    Core of EDP (§3.2, Eqs. 1–4); the paper’s own sensitivity plots show performance drops when α_t is too large, indicating the assumption is only partially true.
  • domain assumption Standard group-relative advantage estimation and PPO-style updates remain stable when advantages from heterogeneous metrics are re-weighted by EMA-driven softmax weights.
    Used in §3.3 (Eqs. 5–9); stability is shown empirically in Fig. 6 but not proved.
invented entities (3)
  • Exploration-Driven Perturbation (EDP) no independent evidence
    purpose: Inject controlled randomness into SFT trajectories so the subsequent RL policy explores task orders and tool choices more broadly.
    Defined by Eqs. 1–4; no independent existence outside this training recipe.
  • Multi-Dimensional Adaptive Reward (MAR) no independent evidence
    purpose: Dynamically re-weight FR/NR metrics via EMA deviation and decoupled advantages to mitigate reward hacking.
    Defined by Eqs. 5–9; the mechanism is paper-specific and validated only by the authors’ ablations.
  • Globally shared Model-Call Pool (MC-Pool) no independent evidence
    purpose: Provide high-concurrency, asynchronous GPU execution of restoration tools during RL rollouts.
    Engineering substrate described in §3.4; necessary for the reported training regime but not an external scientific entity.

pith-pipeline@v1.1.0-grok45 · 20622 in / 3588 out tokens · 60755 ms · 2026-07-13T16:46:15.619574+00:00 · methodology

0 comments
read the original abstract

Vision-language agents that orchestrate specialized tools for image restoration (IR) have emerged as a promising method, yet most existing frameworks operate in a training-free manner. They rely on heuristic task scheduling and exhaustive tool traversal, resulting in sub-optimal restoration paths and prohibitive computational cost. We argue that the core bottleneck lies in the absence of a learned policy to make decision, as a vision-language model cannot efficiently handle degradation-aware task ordering and tool composition. To this end, we propose TIR-Agent, a trainable image restoration agent that performs a direct tool-calling policy through a two-stage training pipeline of supervised fine-tuning (SFT) followed by reinforcement learning (RL). Two key designs underpin effective RL training: (i) a random perturbation strategy applied to the SFT data, which broadens the policy's exploration over task schedules and tool compositions, and (ii) a multi-dimensional adaptive reward mechanism that dynamically re-weights heterogeneous image quality metrics to mitigate reward hacking. To support high-throughput, asynchronous GPU-based tool invocation during training, we further develop a globally shared model-call pool. Experiments on both in-domain and out-of-domain degradations show that TIR-Agent outperforms 12 baselines, including 6 all-in-one models, 3 training-free agents, and 3 proprietary models, and achieves over 2.5$\times$ inference speedup by eliminating redundant tool executions.

discussion (0)

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

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. DiTTo: Scalable Order-aware All-in-One Image Restoration Agent

    cs.CV 2026-05 unverdicted novelty 6.0

    DiTTo reduces optimal restoration trajectory dataset construction from quadratic to linear cost via a simulator and adds order-aware alignment for plug-and-play extensibility to new experts, claiming SOTA quality on m...

  2. EvoIR-Agent: Self-Evolving Image Restoration Agentic System via Experience-Driven Learning

    cs.CV 2026-05 unverdicted novelty 6.0

    EvoIR-Agent formulates experience components into a hierarchical pool with a self-evolving update mechanism to improve performance and efficiency of training-free MLLM image restoration agents over prior paradigms.

  3. EvoIR-Agent: Self-Evolving Image Restoration Agentic System via Experience-Driven Learning

    cs.CV 2026-05 unverdicted novelty 5.0

    EvoIR-Agent introduces a hierarchical experience pool and self-evolving mechanism to improve training-free image restoration agents, claiming significant metric leads and better performance-efficiency balance.

Reference graph

Works this paper leans on

58 extracted references · 21 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Anthropic: Claude opus 4.5 (2025),https://www.anthropic.com/news/claude-opus-4-5

    2025

  2. [2]

    arXiv preprint arXiv:2511.21631 (2025)

    Bai, S., Cai, Y ., Chen, R., et al.: Qwen3-vl technical report. arXiv preprint arXiv:2511.21631 (2025)

    Pith/arXiv arXiv 2025

  3. [3]

    In: CoRL

    Brohan, A., Chebotar, Y ., Finn, C., Hausman, K., Herzog, A., Ho, D., Ibarz, J., Irpan, A., Jang, E., Julian, R., et al.: Do as i can, not as i say: Grounding language in robotic affordances. In: CoRL. pp. 287–318. PMLR (2023)

    2023

  4. [4]

    arXiv preprint arXiv:2410.08688 (2024)

    Cao, J., Meng, D., Cao, X.: Chain-of-restoration: Multi-task image restoration models are zero-shot step-by-step universal image restorers. arXiv preprint arXiv:2410.08688 (2024)

    Pith/arXiv arXiv 2024

  5. [5]

    arXiv preprint arXiv:2410.07095 (2024)

    Chan, J.S., Chowdhury, N., Jaffe, O., Aung, J., Sherburn, D., Mays, E., Starace, G., Liu, K., Maksin, L., Patwardhan, T., Weng, L., M ˛ adry, A.: Mle-bench: Evaluating machine learning agents on machine learning engineering. arXiv preprint arXiv:2410.07095 (2024)

    Pith/arXiv arXiv 2024

  6. [6]

    Advances in Neural Information Processing Systems37, 110643–110666 (2024)

    Chen, H., Li, W., Gu, J., Ren, J., Chen, S., Ye, T., Pei, R., Zhou, K., Song, F., Zhu, L.: Restoreagent: Autonomous image restoration agent via multimodal large language models. Advances in Neural Information Processing Systems37, 110643–110666 (2024)

    2024

  7. [7]

    In: ECCV

    Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: ECCV . pp. 17–33. Springer (2022)

    2022

  8. [8]

    arXiv preprint arXiv:2512.09282 (2025)

    Chen, X., Pan, J., Dong, J., Yang, J., Tang, J.: Foundir-v2: Optimizing pre-training data mixtures for image restoration foundation model. arXiv preprint arXiv:2512.09282 (2025)

    arXiv 2025

  9. [9]

    In: ECCV

    Chen, X., Li, Z., Pu, Y ., Liu, Y ., Zhou, J., Qiao, Y ., Dong, C.: A comparative study of image restoration networks for general backbone network design. In: ECCV . pp. 74–91 (2024)

    2024

  10. [10]

    In: European Conference on Computer Vision

    Conde, M.V ., Geigle, G., Timofte, R.: Instructir: High-quality image restoration following human instructions. In: European Conference on Computer Vision. pp. 1–21. Springer (2024)

    2024

  11. [11]

    In: CVPR

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR. pp. 248–255. IEEE (2009)

    2009

  12. [12]

    IEEE transactions on pattern analysis and machine intelligence38(2), 295–307 (2015)

    Dong, C., Loy, C.C., He, K., Tang, X.: Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence38(2), 295–307 (2015)

    2015

  13. [13]

    arXiv preprint arXiv:2512.10359 (2025)

    Fan, S., Cui, J., Guo, M.H., Yang, S.: Tool-augmented spatiotemporal reasoning for streamlining video question answering task. arXiv preprint arXiv:2512.10359 (2025)

    Pith/arXiv arXiv 2025

  14. [14]

    arXiv preprint arXiv:2504.19678 (2025)

    Ferrag, M.A., Tihanyi, N., Debbah, M.: From llm reasoning to autonomous ai agents: A comprehensive review. arXiv preprint arXiv:2504.19678 (2025)

    Pith/arXiv arXiv 2025

  15. [15]

    Google: Gemini 3 pro: the frontier of vision ai (2025), https://blog.google/innovation-and-ai/ technology/developers-tools/gemini-3-pro-vision/

    2025

  16. [16]

    In: NeurIPS

    Guo, H., Dai, T., Bai, Y ., Chen, B., Ren, X., Zhu, Z., Xia, S.: Parameter efficient adaptation for image restoration with heterogeneous mixture-of-experts. In: NeurIPS. vol. 37, pp. 13522–13547 (2024)

    2024

  17. [17]

    arXiv preprint arXiv:2511.15351 (2025)

    Guo, Y ., Xu, Z., Yao, Z., Lu, Y ., Lin, J., Hu, S., Tang, Z., Wang, H., Chen, R.: Octopus: Agentic multimodal reasoning with six-capability orchestration. arXiv preprint arXiv:2511.15351 (2025)

    arXiv 2025

  18. [18]

    In: CVPR

    Gupta, T., Kembhavi, A.: Visual programming: Compositional visual reasoning without training. In: CVPR. pp. 14953–14962 (2023)

    2023

  19. [19]

    arXiv preprint arXiv:2511.05271 (2025)

    Hong, J., Zhao, C., Zhu, C., Lu, W., Xu, G., Yu, X.: Deepeyesv2: Toward agentic multimodal model. arXiv preprint arXiv:2511.05271 (2025)

    Pith/arXiv arXiv 2025

  20. [20]

    https://helpx.adobe.com/photoshop/using/neural-filters.html (2024)

    Inc., A.: Photoshop neural filters. https://helpx.adobe.com/photoshop/using/neural-filters.html (2024)

    2024

  21. [21]

    arXiv preprint arXiv:2503.09403 (2025)

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

    Pith/arXiv arXiv 2025

  22. [22]

    In: ECCV

    Jiang, Y ., Zhang, Z., Xue, T., Gu, J.: Autodir: Automatic all-in-one image restoration with latent diffusion. In: ECCV . pp. 340–359 (2024)

    2024

  23. [23]

    arXiv preprint arXiv:2401.03379 (2024)

    Kong, X., Dong, C., Zhang, L.: Towards effective multiple-in-one image restoration: A sequential and prompt learning strategy. arXiv preprint arXiv:2401.03379 (2024)

    Pith/arXiv arXiv 2024

  24. [24]

    Labs, T.: Topaz photo ai.https://www.topazlabs.com/topaz-photo-ai(2024)

    2024

  25. [25]

    arXiv preprint arXiv:2503.10120 (2025)

    Li, B., Li, X., Lu, Y ., Chen, Z.: Hybrid agents for image restoration. arXiv preprint arXiv:2503.10120 (2025)

    Pith/arXiv arXiv 2025

  26. [26]

    In: CVPR

    Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR. pp. 17452–17462 (2022) 12 TIR-Agent: Training an Explorative and Efficient Agent for Image Restoration

    2022

  27. [27]

    In: ICCV

    Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. In: ICCV . pp. 1833–1844 (2021)

    2021

  28. [28]

    In: CVPR-Workshop

    Lim, B., Son, S., Kim, H., Nah, S., Mu Lee, K.: Enhanced deep residual networks for single image super-resolution. In: CVPR-Workshop. pp. 136–144 (2017)

    2017

  29. [29]

    arXiv preprint arXiv:2601.05242 (2026)

    Liu, S.Y ., Dong, X., Lu, X., Diao, S., Belcak, P., Liu, M., Chen, M.H., Yin, H., Wang, Y .C.F., Cheng, K.T., et al.: Gdpo: Group reward-decoupled normalization policy optimization for multi-reward rl optimization. arXiv preprint arXiv:2601.05242 (2026)

    Pith/arXiv arXiv 2026

  30. [30]

    arXiv preprint arXiv:2512.18599 (2025)

    Lu, J., Wu, Y ., Zhao, Z., Wang, H., Jimenez, F., Majeedi, A., Fu, Y .: Simplecall: A lightweight image restoration agent in label-free environments with mllm perceptual feedback. arXiv preprint arXiv:2512.18599 (2025)

    Pith/arXiv arXiv 2025

  31. [31]

    arXiv preprint arXiv:2310.01018 (2023)

    Luo, Z., Gustafsson, F.K., Zhao, Z., Sjölund, J., Schön, T.B.: Controlling vision-language models for multi-task image restoration. arXiv preprint arXiv:2310.01018 (2023)

    Pith/arXiv arXiv 2023

  32. [32]

    arXiv preprint arXiv:2306.13653 (2023)

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

    Pith/arXiv arXiv 2023

  33. [33]

    OpenAI: Introducing gpt-5.2 (2025),https://openai.com/index/introducing-gpt-5-2/

    2025

  34. [34]

    In: CVPR

    Park, D., Lee, B.H., Chun, S.Y .: All-in-one image restoration for unknown degradations using adaptive discrimi- native filters for specific degradations. In: CVPR. pp. 5815–5824. IEEE (2023)

    2023

  35. [35]

    Advances in neural information processing systems36, 71275–71293 (2023)

    Potlapalli, V ., Zamir, S.W., Khan, S.H., Shahbaz Khan, F.: Promptir: Prompting for all-in-one image restoration. Advances in neural information processing systems36, 71275–71293 (2023)

    2023

  36. [36]

    arXiv preprint arXiv:2504.04903 (2025)

    Pu, Y ., Zhuo, L., Zhu, K., Xie, L., Zhang, W., Chen, X., Gao, P., Qiao, Y ., Dong, C., Liu, Y .: Lumina-omnilv: A unified multimodal framework for general low-level vision. arXiv preprint arXiv:2504.04903 (2025)

    Pith/arXiv arXiv 2025

  37. [37]

    Schuhmann, C.: Improved aesthetic predictor (2022), https://github.com/christophschuhmann/ improved-aesthetic-predictor

    2022

  38. [38]

    Advances in Neural Information Processing Systems36, 38154–38180 (2023)

    Shen, Y ., Song, K., Tan, X., Li, D., Lu, W., Zhuang, Y .: Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face. Advances in Neural Information Processing Systems36, 38154–38180 (2023)

    2023

  39. [39]

    In: NeurIPS (2025)

    Su, A., Wang, H., Ren, W., Lin, F., Chen, W.: Pixel reasoner: Incentivizing pixel space reasoning via curiosity- driven reinforcement learning. In: NeurIPS (2025)

    2025

  40. [40]

    In: ICCV

    Surís, D., Menon, S., V ondrick, C.: Vipergpt: Visual inference via python execution for reasoning. In: ICCV . pp. 11888–11898 (2023)

    2023

  41. [41]

    arXiv preprint arXiv:2512.17532 (2025)

    Tang, J., Chen, J., Wei, W., Xu, X., Liu, R., Wu, X., Xie, Q., Wu, J., Zhang, L., Chen, Q.: Robust-r1: Degradation- aware reasoning for robust visual understanding. arXiv preprint arXiv:2512.17532 (2025)

    arXiv 2025

  42. [42]

    5: Visual agentic intelligence

    Team, K., Bai, T., Bai, Y ., Bao, Y ., Cai, S., Cao, Y ., Charles, Y ., Che, H., Chen, C., Chen, G., et al.: Kimi k2. 5: Visual agentic intelligence. arXiv preprint arXiv:2602.02276 (2026)

    Pith/arXiv arXiv 2026

  43. [43]

    arXiv preprint arXiv:2512.16918 (2025)

    Wang, C., Feng, K., Chen, D., Wang, Z., Li, Z., Gao, S., Meng, M., Zhou, X., Zhang, M., Shang, Y ., et al.: Adatooler-v: Adaptive tool-use for images and videos. arXiv preprint arXiv:2512.16918 (2025)

    Pith/arXiv arXiv 2025

  44. [44]

    Wang, X., Xie, L., Yu, K., Chan, K.C., Loy, C.C., Dong, C.: Basicsr: Open source image and video restoration toolbox.https://github.com/XPixelGroup/BasicSR(2022)

    2022

  45. [45]

    In: ECCV-Workshop

    Wang, X., Yu, K., Wu, S., Gu, J., Liu, Y ., Dong, C., Qiao, Y ., Change Loy, C.: Esrgan: Enhanced super-resolution generative adversarial networks. In: ECCV-Workshop. pp. 0–0 (2018)

    2018

  46. [46]

    In: CVPR

    Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: CVPR. pp. 17683–17693 (2022)

    2022

  47. [47]

    arXiv preprint arXiv:2303.04671 (2023)

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

    Pith/arXiv arXiv 2023

  48. [48]

    In: NeurIPS

    Yang, L., Kang, B., Huang, Z., Zhao, Z., Xu, X., Feng, J., Zhao, H.: Depth anything v2. In: NeurIPS. pp. 21875–21911 (2024)

    2024

  49. [49]

    IEEE transactions on image processing33, 5408–5423 (2024)

    Yao, M., Xu, R., Guan, Y ., Huang, J., Xiong, Z.: Neural degradation representation learning for all-in-one image restoration. IEEE transactions on image processing33, 5408–5423 (2024)

    2024

  50. [50]

    IEEE Transactions on Image Processing (2025)

    You, Z., Gu, J., Cai, X., Li, Z., Zhu, K., Dong, C., Xue, T.: Enhancing descriptive image quality assessment with a large-scale multi-modal dataset. IEEE Transactions on Image Processing (2025)

    2025

  51. [51]

    In: ECCV

    You, Z., Li, Z., Gu, J., Yin, Z., Xue, T., Dong, C.: Depicting beyond scores: Advancing image quality assessment through multi-modal language models. In: ECCV . pp. 259–276. Springer (2024) 13 TIR-Agent: Training an Explorative and Efficient Agent for Image Restoration

    2024

  52. [52]

    In: CVPR

    Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR. pp. 5728–5739 (2022)

    2022

  53. [53]

    arXiv preprint arXiv:2508.04416 (2025)

    Zhang, H., Gu, X., Li, J., Ma, C., Bai, S., Zhang, C., Zhang, B., Zhou, Z., He, D., Tang, Y .: Thinking with videos: Multimodal tool-augmented reinforcement learning for long video reasoning. arXiv preprint arXiv:2508.04416 (2025)

    Pith/arXiv arXiv 2025

  54. [54]

    IEEE transactions on image processing26(7), 3142–3155 (2017)

    Zhang, K., Zuo, W., Chen, Y ., Meng, D., Zhang, L.: Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing26(7), 3142–3155 (2017)

    2017

  55. [55]

    thinking with images

    Zheng, Z., Yang, M., Hong, J., Zhao, C., Xu, G., Yang, L., Shen, C., Yu, X.: Deepeyes: Incentivizing" thinking with images" via reinforcement learning. arXiv preprint arXiv:2505.14362 (2025)

    Pith/arXiv arXiv 2025

  56. [56]

    arXiv preprint arXiv:2504.07148 (2025)

    Zhou, Y ., Cao, J., Zhang, Z., Wen, F., Jiang, Y ., Jia, J., Liu, X., Min, X., Zhai, G.: Q-agent: Quality-driven chain-of- thought image restoration agent through robust multimodal large language model. arXiv preprint arXiv:2504.07148 (2025)

    Pith/arXiv arXiv 2025

  57. [57]

    arXiv preprint arXiv:2410.17809 (2024)

    Zhu, K., Gu, J., You, Z., Qiao, Y ., Dong, C.: An intelligent agentic system for complex image restoration problems. arXiv preprint arXiv:2410.17809 (2024)

    Pith/arXiv arXiv 2024

  58. [58]

    In: NeurIPS (2025) 14

    Zuo, Y ., Zheng, Q., Wu, M., Jiang, X., Li, R., Wang, J.: 4kagent: Agentic any image to 4k super-resolution. In: NeurIPS (2025) 14

    2025