Pith. sign in

REVIEW 4 major objections 5 minor 72 references

A self-evolving causal memory graph lets image-restoration agents learn from every action and reach 35.55 dB average PSNR on combined rain, snow, haze, blur, and noise.

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 · deepseek-v4-flash

2026-08-01 08:22 UTC pith:6BG6LUW4

load-bearing objection Novel memory-graph idea for restoration agents, but the empirical claims are conditional on a missing evaluation protocol and a circular memory reward. the 4 major comments →

arxiv 2607.21125 v1 pith:6BG6LUW4 submitted 2026-07-23 cs.CV

Causal-AgentIR: Self-Evolving Causal Memory for Adaptive Image Restoration Agents

Hu Gao , Yulong Chen , Lizhuang Ma This is my paper
classification cs.CV
keywords image restorationmulti-agent systemcausal memory graphself-evolving memorydegradation-aware planningtool selectionexecution orderinguser preference
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.

The paper proposes Causal-AgentIR, a hierarchical multi-agent system for image restoration that records every executed restoration step—degradation, region, tool, action, quality change, cost, user preference—in a structured causal memory graph and updates that graph after each action. Its central claim is that this self-evolving memory, rather than any new restoration network, is what lets the system choose better tools and better execution orders over time. A sympathetic reader would take the paper as establishing that image restoration can be framed as a lifelong learning problem: with the same frozen tool library, the agent improves because it remembers which operations help or harm under which conditions. The reported results put the all-in-one average at 35.55 dB PSNR, ahead of every compared agent and model, with the ablation attributing 0.40 dB specifically to memory evolution and 0.26 dB to multi-agent collaboration.

Core claim

Causal-AgentIR claims that restoration experience is better stored as a graph than as text or weights. Nodes stand for degradations, image regions, restoration tools, actions, quality-change measurements, computational cost, and user preferences; edges carry a contextual condition, an estimated effect on restoration quality, and a confidence score. During inference the causal-memory agent retrieves relevant subgraphs and scores candidate plans by combining edge confidences and effects across causal paths, so the planner can weigh whether to apply a tool globally or regionally and in what order. After each action, a memory curator evaluates the observed quality change and decides to add, upda

What carries the argument

The self-evolving causal memory graph: a directed graph whose nodes encode degradation patterns, image regions, restoration tools, actions, quality changes, computational costs, and user preferences, and whose edges store a contextual condition, an estimated effect on restoration quality, and a confidence score. The graph carries the argument because it makes action–outcome relations retrievable, combinable, and updatable: a query over the current image state and user instruction retrieves relevant subgraphs, path-level causal scores turn those subgraphs into expected utility for candidate plans, and a memory curator revises edge weights and confidences after each executed action. The paper

Load-bearing premise

The reported gains stand only if the memory graph is populated from prior experience and not from the test images being scored; the paper updates memory after every executed action but never states whether memory is reset or frozen across the benchmarks it calls zero-shot.

What would settle it

Run the all-in-one benchmark twice with identical agents and tool library—once with the memory graph frozen at its initial state and once with it evolving during evaluation. If the evolving version keeps its reported ~0.40 dB advantage, self-evolution is doing real work; if the advantage disappears, the gains come from adapting to test images.

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

If this is right

  • With frozen restoration tools, the reported coordination gains (0.73 dB over the no-memory, no-collaboration baseline) come entirely from planning, memory-guided ordering, and regional application.
  • Self-evolving causal memory alone contributes 0.40 dB and cuts rollback rate from 0.38 to 0.22, implying the agent avoids harmful operations before executing them.
  • The best all-in-one restoration result reported, 35.55 dB average PSNR, surpasses every compared model and agent on the combined rain, snow, haze, blur, and noise benchmark.
  • Real-world zero-shot generalization improves on no-reference metrics for rain, haze, and noise, suggesting memory-based planning transfers to degradations outside the training distribution.
  • Recorded user preferences raise preference-alignment rate to 0.90, so the same memory mechanism adapts restoration behavior to individual users.

Where Pith is reading between the lines

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

  • Editorial inference: because the graph format is tool-agnostic, the same self-evolving memory mechanism could coordinate tools in other vision agent tasks—video restoration, enhancement, or editing—without changing the graph design.
  • Editorial inference: the paper's 'causal' edges come from observed trajectories, not controlled interventions; a stricter causal test would randomize or counterfactually reorder tool sequences to check whether the learned effect estimates hold up.
  • Editorial inference: the reported zero-shot results would be cleaner if memory were reset or frozen across benchmarks; without that control, gains could partly reflect test-time adaptation rather than transferable knowledge.

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 / 5 minor

Summary. The paper proposes Causal-AgentIR, a hierarchical multi-agent framework for image restoration in which restoration experience is organized as a self-evolving causal memory graph. Nodes represent degradations, regions, tools, actions, quality changes, costs, and user preferences; edges carry contextual conditions, causal-effect weights, and confidence scores. A memory-curator agent adds, updates, merges, reinforces, ignores, or discards edges after each executed action, while a planner, degradation agent, tool agents, causal-memory agent, and critic collaborate to select restoration actions. Experiments cover task-specific, all-in-one, real-world, and mixed-degradation settings. The paper reports best or near-best PSNR/SSIM in many benchmarks and attributes the improvement to multi-agent collaboration (+0.26 dB) and self-evolving causal memory (+0.40 dB) in Table XII.

Significance. If the reported results are reproducible and the memory protocol is valid, the structured causal memory graph is a meaningful step beyond static tool descriptions and unstructured textual summaries in agentic image restoration. The explicit modeling of confidence, effect magnitude, and selective forgetting is a promising design. The paper also provides extensive benchmark comparisons, ablations, and a user study. However, the manuscript currently omits the training procedure for the learnable memory policy, the values of more than a dozen hyperparameters, and any statement about whether the memory graph is reset or frozen during evaluation. These omissions are load-bearing for the central claim that the gains come from transferable causal memory rather than test-time adaptation or reward self-consistency.

major comments (4)
  1. [§IV-E / §III-E, Eq. (21)] The real-world results are called 'zero-shot' in §IV-E, but the paper never states whether the memory graph G is reset per image, frozen after a training phase, or carried across test images. Since §III-E updates memory after every executed action, a persistent graph lets later test images benefit from edges learned on earlier test images, while a per-image reset turns the method into online adaptation that the baselines do not perform. Either way, the +0.40 dB memory gain in Table XII does not isolate transferable causal knowledge unless the protocol is specified and a no-adaptation variant is reported.
  2. [§III-E, Eq. (31)] The memory reward R_mem_t = U(Π_{t+1}|G_{t+1}) − U(Π_{t+1}|G_t) is computed from the same utility U that uses edge weights w and confidences γ that the curator is updating. The update rules in Eqs. (24)–(25) push w and γ toward the currently observed ΔQ_t, so R_mem_t can be positive by construction even if future restoration quality is unchanged. This reward therefore measures self-consistency of the memory update, not an independent improvement in future planning. Please redefine R_mem using held-out planner outcomes or a separate validation reward.
  3. [§III-E, Eqs. (21)–(31)] The memory-operation policy π_m is called 'learnable,' but the manuscript provides no training objective, dataset, optimization procedure, or loss function for it; the equations describe only the operation taxonomy and reward. Furthermore, the values of α, ρ, β1–β5, τ_Q, τ_δ, τ_γ, τ_w, λ_s, λ_γ, λ_w, λ_C, γ0, and η1 are not reported. Without these details, the experiments are not reproducible, and it is unclear whether the policy is actually trained or is a hand-crafted rule. Please provide the full training protocol and all hyperparameter values.
  4. [§III-D, Eqs. (14) and (16)] The retrieval relevance in Eq. (14) depends on a similarity function Sim(ψ_t, z_e) that is never defined, and on edge confidence γ_e and effect |w_e| without specifying how these are normalized. Similarly, Eq. (16) aggregates over causal paths Ω(Π_t, g_t) without defining how paths are enumerated, their maximum length, or how the product-of-γ and sum-of-w in Eq. (17) are normalized across plans with different path counts. These details are necessary to determine whether the planner genuinely uses multi-hop causal reasoning or effectively uses a single-edge heuristic.
minor comments (5)
  1. [§IV-F] The section heading has a stray space: 'F . Agent Efficiency...' Please fix the formatting.
  2. [§IV-A] The abbreviations R+S+H+B+N and H+R+N are used without being spelled out. Please define them at first use.
  3. [§III-E, Eq. (19)] The components of ΔQ_t (fidelity, perceptual, objective, artifact, cost, preference) have different units and scales, but no normalization protocol is given. Without this, the weights β1–β5 and the threshold τ_Q are not meaningful.
  4. [§IV-I, Fig. 6] Figure 6 refers to 'ChatGPT 4o' and 'the all-in-one method' without specifying which all-in-one method was used. The user study is described as small-scale; please report the number of comparisons per condition and any variability/confidence measure.
  5. [General] Releasing code, trained memory policies, and the exact memory-initialization/reset script would substantially increase the credibility of the empirical claims.

Circularity Check

2 steps flagged

Memory-evolution reward is self-confirming by construction; zero-shot evaluation lacks a stated memory-reset protocol.

specific steps
  1. self definitional [Section III-E, Eqs. (16)-(17), (23), (24), (30), (31)]
    "R_t = ΔQ_t + η_1 R_mem_t (30); R_mem_t = U(Π_{t+1}|G_{t+1}) − U(Π_{t+1}|G_t) (31); w^{t+1}_{ij} = α w^t_{ij} + (1−α)ΔQ_t (24); e_new = (v_i, v_j, z_new, ΔQ_t, γ_0) (23)"

    By Eq. (23)/(24), the memory edge weight w_e is set or updated with the observed ΔQ_t. By Eqs. (16)-(17), the utility U is a path sum over these same w_e and γ_e values. Therefore, whenever ΔQ_t is positive, the update increases w_e (or adds a positive edge) and 'Reinforce' raises γ_e, so U(G_{t+1}) > U(G_t) mechanically. R_mem thus rewards the curator for recording the very observation that produced the reward, rather than measuring whether future planning improved. The 'predicted utility' is a smoothed echo of the quality signal used to write the memory, so the memory-evolution objective is self-confirming by construction.

  2. fitted input called prediction [Section III-E Eq. (21); Section IV-E1 'Zero-shot generalization in real-world scenes']
    "After executing action a_t, the critic and memory curator agents evaluate the transition from I_t to I_{t+1} and update the causal memory accordingly. ... We further evaluate the zero-shot generalization capability of different methods ... on real-world benchmarks, including RealRain-1k-L, RTTS, and SIDD."

    The paper never states whether the memory graph G is reset per image, frozen after a training phase, or carried across the benchmark, nor whether G is initialized from experience on the same degradation types used in evaluation. Because Eq. (21) applies the curator update during evaluation, a persistent G lets later test images inherit edges learned from earlier test images, making the 'zero-shot' results an evaluation-order artifact; a per-image reset instead turns the method into per-image online adaptation that baselines do not perform. In either reading, the reported real-world gains are at least partially fitted to the evaluation set rather than predicted from fixed transferable knowledge, so the zero-shot claim is unsupported as written.

full rationale

The paper's main empirical claims rest on fixed restoration tools and external PSNR measurements, so the overall comparison is not a pure derivation-to-input tautology, and there is no load-bearing self-citation chain or imported uniqueness theorem. However, the formal memory-evolution reward (Eq. 31) is self-referential: it evaluates the change in a utility function computed from the same edge weights and confidences that the curator just updated using ΔQ_t, so positive observations mechanically produce positive memory rewards. In addition, the paper calls its real-world evaluation 'zero-shot' but does not specify whether memory is reset or frozen across the benchmark; as written, the protocol permits memory to be fitted to the evaluation data. These issues are partial circularity rather than total collapse: the ablation attributing +0.40 dB to memory would be credible only with an explicit held-out memory protocol that the text does not provide.

Axiom & Free-Parameter Ledger

9 free parameters · 4 axioms · 2 invented entities

The central claim rests on many unstated or untuned parameters and on the domain assumption that frozen tools plus memory retrieval yields transferable knowledge. No independent evidence is provided for the memory policy; the memory's improvement is measured only within the agent's own utility estimate.

free parameters (9)
  • α (momentum coefficient, Eq. 24)
    Controls how fast old causal effect w_ij is replaced by new ΔQ_t; no value given.
  • ρ (reinforcement rate, Eq. 25)
    Controls confidence increase for consistent evidence; no value given.
  • β1..β5 (quality-change weights, Eq. 19)
    Balance fidelity, perceptual, objective, artifact, cost, and preference terms in ΔQ_t; no values given.
  • τ_Q, τ_δ (ignore thresholds, Eq. 28)
    Determine which experiences are ignored as uninformative; no values given.
  • τ_γ, τ_w (discard thresholds, Eq. 29)
    Determine when low-confidence negative edges are removed; no values given.
  • λ_s, λ_γ, λ_w (retrieval weights, Eq. 14)
    Directly control how much context similarity, confidence, and effect magnitude matter in retrieval; no values or sensitivity analysis given.
  • λ_C (cost weight, Eq. 16)
    Balances computational cost against expected causal utility; no value given.
  • γ0 (initial edge confidence, Eq. 23)
    Sets the starting confidence of newly added causal edges; no value given.
  • η1 (memory reward weight, Eq. 30)
    Balances immediate restoration quality against estimated future benefit of memory updates; no value given.
axioms (4)
  • domain assumption The memory graph can be updated online from observed transitions and this improves future decisions (Sections III-D/E).
    The entire advantage claim depends on ΔQ_t-derived edge weights w_e generalizing to new images rather than merely memorizing the current benchmark. No protocol separates training vs. test memory state.
  • domain assumption Causal edges can be estimated from single-tool executions without counterfactual controls or noise modeling (Section III-D).
    The paper explicitly says 'causal' is an empirical interventional relation induced by actual restoration operations, not a structural causal model. This weaker assumption is load-bearing because all planning uses these edge weights.
  • domain assumption The frozen restoration tool library contains tools adequate for every degradation encountered, and the quality metrics used capture what users want (Section IV-A).
    If the tool library lacks a needed tool or metrics misalign with preferences, memory cannot compensate. The method only selects/orders existing tools; it never modifies them.
  • standard math Maximum expected utility with additive rewards is a valid objective for restoration trajectories (Eq. 4).
    Standard MDP-style formulation; not controversial by itself.
invented entities (2)
  • Causal memory graph G=(V,E) with causal-effect edges e_ij=(v_i,v_j,z_ij,w_ij,γ_ij) no independent evidence
    purpose: Encodes restoration experience for structured retrieval and planning (Eq. 11).
    The graph's causal validity is never checked against an external ground truth; improvements are attributed to it only via the agent's own utility estimates and downstream metrics.
  • Memory-operation policy π_m(o|ξ_t,g_t) no independent evidence
    purpose: Selects add/update/merge/reinforce/ignore/discard for each transition (Eq. 21).
    Called 'learnable' but no training data, loss, or learned parameters are disclosed; no independent evidence that the learned policy exists beyond the prose.

pith-pipeline@v1.3.0-alltime-deepseek · 20138 in / 14421 out tokens · 120011 ms · 2026-08-01T08:22:28.083069+00:00 · methodology

0 comments
read the original abstract

Image restoration agents have recently emerged as a flexible paradigm for handling diverse and unpredictable degradations in real-world scenarios. Existing agents typically formulate restoration as a tool-using process, where the agent perceives degradations, searches candidate tools, executes restoration operations, and revises the plan through reflection or rollback. However, their knowledge is often stored as static tool descriptions, manually defined degradation priors, or unstructured textual summaries, which limits the accumulation, verification, revision, and forgetting of restoration knowledge over long-term experience. In this paper, we propose Causal-AgentIR, a hierarchical multi-agent framework with self-evolving causal memory for collective image restoration intelligence. Instead of representing restoration experience as isolated textual records, Causal-AgentIR organizes degradation patterns, image regions, restoration tools, actions, quality changes, and user preferences into a structured causal memory graph. This graph supports graph-based retrieval and multi-hop causal reasoning, enabling agents to infer how specific restoration operations or tool sequences affect restoration quality under different degradation conditions. The framework further organizes multiple agents into a collaborative system, including planning, degradation analysis, tool expertise, causal memory reasoning, outcome critique, and memory curation. Through this design, restoration experience can be added, updated, merged, reinforced, ignored, or discarded according to observed quality changes and feedback, allowing the agent to maintain reliable and transferable restoration knowledge. Extensive experiments demonstrate the effectiveness of the proposed framework.

Figures

Figures reproduced from arXiv: 2607.21125 by Hu Gao, Lizhuang Ma, Yulong Chen.

Figure 1
Figure 1. Figure 1: Comparison between existing image restoration paradigms and the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed Causal-AgentIR framework. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of different restoration methods. The fourth column presents the results of the corresponding task-specific model, while the [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Multi-degradation image restoration results across various degradation combinations. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: User study results following the setting of IAMAgent [26]. (a) Causal [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 5
Figure 5. Figure 5: Case studies of Causal-AgentIR. ratio (PAR) of 0.76. Introducing the multi-agent mechanism improves PSNR to 35.08 dB, yielding a gain of 0.26 dB. It also increases SR from 0.82 to 0.88 and PAR from 0.76 to 0.82, indicating that collaborative degradation analysis, planning, tool execution, and critique improve restoration reliability and user-preference alignment. When only self-evolving causal memory is us… view at source ↗

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

72 extracted references · 4 linked inside Pith

  1. [1]

    Pixel2pixel: A pixelwise approach for zero-shot single image denoising,

    Q. Ma, J. Jiang, X. Zhou, P. Liang, X. Liu, and J. Ma, “Pixel2pixel: A pixelwise approach for zero-shot single image denoising,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 47, no. 6, pp. 4614–4629, 2025

    2025

  2. [2]

    Learning enriched features via selective state spaces model for efficient image deblurring,

    H. Gao, B. Ma, Y . Zhang, J. Yang, J. Yang, and D. Dang, “Learning enriched features via selective state spaces model for efficient image deblurring,” inProceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 710–718

    2024

  3. [3]

    Efficientderain+: Learning uncertainty-aware filtering via rainmix augmentation for high-efficiency deraining,

    Q. Guo, H. Qi, J. Sun, F. Juefei-Xu, L. Ma, D. Lin, W. Feng, and S. Wang, “Efficientderain+: Learning uncertainty-aware filtering via rainmix augmentation for high-efficiency deraining,”International Journal of Computer Vision, vol. 133, no. 4, pp. 2111–2135, 2025

    2025

  4. [4]

    Prior-guided hierarchical harmonization net- work for efficient image dehazing,

    X. Su, S. Li, Y . Cui, M. Cao, Y . Zhang, Z. Chen, Z. Wu, Z. Wang, Y . Zhang, and X. Yuan, “Prior-guided hierarchical harmonization net- work for efficient image dehazing,” inProceedings of the AAAI Confer- ence on Artificial Intelligence, vol. 39, no. 7, 2025, pp. 7042–7050

    2025

  5. [5]

    Learn- ing optimal combination patterns for lightweight stereo image super- resolution,

    H. Gao, J. Yang, Y . Zhang, J. Yang, B. Ma, and D. Dang, “Learn- ing optimal combination patterns for lightweight stereo image super- resolution,” inProceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 5566–5574

    2024

  6. [6]

    Image restoration via frequency selection,

    Y . Cui, W. Ren, X. Cao, and A. Knoll, “Image restoration via frequency selection,”TPAMI, pp. 1–16, 2023

    2023

  7. [7]

    Mixed hierarchy network for image restoration,

    H. Gao, Y . Zhang, J. Yang, and D. Dang, “Mixed hierarchy network for image restoration,”Pattern Recognition, vol. 161, p. 111313, 2025

    2025

  8. [8]

    Selective frequency network for image restoration,

    Y . Cui, Y . Tao, Z. Bing, W. Ren, X. Gao, X. Cao, K. Huang, and A. Knoll, “Selective frequency network for image restoration,” inICLR, 2023

    2023

  9. [9]

    Acl: Activating capability of linear attention for image restoration,

    Y . Gu, Y . Meng, J. Ji, and X. Sun, “Acl: Activating capability of linear attention for image restoration,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 17 913–17 923

    2025

  10. [10]

    Starir: Convolutional image restoration with spatial-frequency fusion,

    Y . Cui, S. W. Zamir, M.-H. Yang, A. Knoll, F. S. Khan, and S. Khan, “Starir: Convolutional image restoration with spatial-frequency fusion,” IEEE Transactions on Pattern Analysis and Machine Intelligence, pp. 1–18, 2026

    2026

  11. [11]

    Mambairv2: Attentive state space restoration,

    H. Guo, Y . Guo, Y . Zha, Y . Zhang, W. Li, T. Dai, S.-T. Xia, and Y . Li, “Mambairv2: Attentive state space restoration,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 28 124– 28 133

    2025

  12. [12]

    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,”Advances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  13. [13]

    Prompt- based ingredient-oriented all-in-one image restoration,

    H. Gao, J. Yang, Y . Zhang, N. Wang, J. Yang, and D. Dang, “Prompt- based ingredient-oriented all-in-one image restoration,”IEEE Transac- tions on Circuits and Systems for Video Technology, vol. 34, no. 10, pp. 9458–9471, 2024

    2024

  14. [14]

    Learning continuous wasser- stein barycenter space for generalized all-in-one image restoration,

    X. Tang, X. He, J. Xu, X. Gu, and J. Sun, “Learning continuous wasser- stein barycenter space for generalized all-in-one image restoration,” IEEE Transactions on Pattern Analysis and Machine Intelligence, pp. 1–16, 2026

    2026

  15. [15]

    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,” inComputer Vision – ECCV 2024: 18th European Conference, Milan, Italy, September 29–October 4, 2024, Proceedings, Part XL, 2024, p. 340–359

    2024

  16. [16]

    Vision-language gradient descent-driven all-in-one deep unfolding networks,

    H. Zeng, X. Wang, Y . Chen, J. Su, and J. Liu, “Vision-language gradient descent-driven all-in-one deep unfolding networks,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2025, pp. 7524–7533

    2025

  17. [17]

    Language mod- els 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 mod- els are few-shot learners,”Advances in neural information processing systems, vol. 33, pp. 1877–1901, 2020

    1901

  18. [18]

    Code llama: Open foundation models for code,

    B. Roziere, J. Gehring, F. Gloeckle, S. Sootla, I. Gat, X. E. Tan, Y . Adi, J. Liu, R. Sauvestre, T. Remezet al., “Code llama: Open foundation models for code,”arXiv preprint arXiv:2308.12950, 2023

    Pith/arXiv arXiv 2023

  19. [19]

    Llama 2: Open foundation and fine-tuned chat models,

    H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y . Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosaleet al., “Llama 2: Open foundation and fine-tuned chat models,”arXiv preprint arXiv:2307.09288, 2023

    Pith/arXiv arXiv 2023

  20. [20]

    Large multimodal agents: a survey,

    J. Xie, Z. Chen, R. Zhang, and G. Li, “Large multimodal agents: a survey,”Visual Intelligence, vol. 3, no. 1, p. 24, 2025

    2025

  21. [21]

    Magma: A foundation model for multimodal ai agents,

    J. Yang, R. Tan, Q. Wu, R. Zheng, B. Peng, Y . Liang, Y . Gu, M. Cai, S. Ye, J. Janget al., “Magma: A foundation model for multimodal ai agents,” inProceedings of the computer vision and pattern recognition conference, 2025, pp. 14 203–14 214

    2025

  22. [22]

    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,” inProceedings of the 38th International Conference on Neural Information Processing Systems, 2024

    2024

  23. [23]

    Multi-agent image restoration,

    X. Jiang, G. Li, B. Chen, and J. Zhang, “Multi-agent image restoration,” International Journal of Computer Vision, vol. 134, no. 5, p. 205, 2026

    2026

  24. [24]

    An intelligent agen- tic system for complex image restoration problems,

    K. Zhu, J. Gu, Z. You, Y . Qiao, and C. Dong, “An intelligent agen- tic system for complex image restoration problems,” inInternational Conference on Learning Representations, vol. 2025, 2025, pp. 57 985– 58 013

    2025

  25. [25]

    Hybrid agents for image restoration,

    B. Li, X. Li, Y . Lu, and Z. Chen, “Hybrid agents for image restoration,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 22 636–22 647

    2026

  26. [26]

    Iamagent: Towards an interactive and adaptive multi-agent system for image restoration,

    Y . Wei, Y . Zhang, H. Zheng, J. Ren, X. Xu, Z. Shi, Z. Zhang, and M. Wang, “Iamagent: Towards an interactive and adaptive multi-agent system for image restoration,”IEEE Transactions on Image Processing, 2026

    2026

  27. [27]

    Restore-r1: Efficient image restoration agents via reinforcement learn- ing with multimodal llm perceptual feedback,

    J. Lu, Y . Wu, Z. Zhao, H. Wang, F. Jimenez, A. Majeedi, and Y . Fu, “Restore-r1: Efficient image restoration agents via reinforcement learn- ing with multimodal llm perceptual feedback,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 8629–8639

    2026

  28. [28]

    Single image haze removal using dark channel prior

    K. He, J. Sun, and X. Tang, “Single image haze removal using dark channel prior.”TPAMI, 2011. 13

    2011

  29. [29]

    Imu-assisted accurate blur kernel re- estimation in non-uniform camera shake deblurring,

    J. Rong, H. Huang, and J. Li, “Imu-assisted accurate blur kernel re- estimation in non-uniform camera shake deblurring,”IEEE Transactions on Image Processing, vol. 33, pp. 3823–3838, 2024

    2024

  30. [30]

    Learning physics-informed noise models from dark frames for low-light raw image denoising,

    H. Feng, L. Wang, Y . Huang, Y . Wang, L. Zhu, and H. Huang, “Learning physics-informed noise models from dark frames for low-light raw image denoising,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 48, no. 4, pp. 3952–3969, 2026

    2026

  31. [31]

    Frequency domain task-adaptive network for restoring images with combined degradations,

    H. Gao, B. Ma, Y . Zhang, J. Yang, J. Yang, and D. Dang, “Frequency domain task-adaptive network for restoring images with combined degradations,”Pattern Recognition, vol. 158, p. 111057, 2025

    2025

  32. [32]

    Xyscannet: An interpretable state space model for perceptual image deblurring,

    H. Liu, C. Liu, J. Xu, P. Jiang, and M. Lu, “Xyscannet: An interpretable state space model for perceptual image deblurring,” inProceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2025, pp. 779–789

    2025

  33. [33]

    Unpaired photo-realistic image deraining with energy-informed diffusion model,

    Y . Wen, T. Gao, and T. Chen, “Unpaired photo-realistic image deraining with energy-informed diffusion model,” inProceedings of the 32nd ACM International Conference on Multimedia, 2024, p. 360–369

    2024

  34. [34]

    Unmix- ing diffusion for self-supervised hyperspectral image denoising,

    H. Zeng, J. Cao, K. Zhang, Y . Chen, H. Luong, and W. Philips, “Unmix- ing diffusion for self-supervised hyperspectral image denoising,” in2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 27 820–27 830

    2024

  35. [35]

    Neural degradation representation learning for all-in-one image restoration,

    M. Yao, R. Xu, Y . Guan, J. Huang, and Z. Xiong, “Neural degradation representation learning for all-in-one image restoration,”IEEE Transac- tions on Image Processing, vol. 33, pp. 5408–5423, 2024

    2024

  36. [36]

    Visual-instructed degradation diffusion for all-in-one image restoration,

    W. Luo, H. Qin, Z. Chen, L. Wang, D. Zheng, Y . Li, Y . Liu, B. Li, and W. Hu, “Visual-instructed degradation diffusion for all-in-one image restoration,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2025, pp. 12 764–12 777

    2025

  37. [37]

    AdaIR: Adaptive all-in-one image restoration via frequency mining and modulation,

    Y . Cui, S. W. Zamir, S. Khan, A. Knoll, M. Shah, and F. S. Khan, “AdaIR: Adaptive all-in-one image restoration via frequency mining and modulation,” inThe Thirteenth International Conference on Learning Representations, 2025

    2025

  38. [38]

    Perceive-ir: Learning to perceive degradation better for all-in-one image restoration,

    X. Zhang, J. Ma, G. Wang, Q. Zhang, H. Zhang, and L. Zhang, “Perceive-ir: Learning to perceive degradation better for all-in-one image restoration,”IEEE Transactions on Image Processing, pp. 1–1, 2025

    2025

  39. [39]

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

    J. Li, D. Li, S. Savarese, and S. Hoi, “Blip-2: Bootstrapping language- image pre-training with frozen image encoders and large language models,” inInternational conference on machine learning, 2023, pp. 19 730–19 742

    2023

  40. [40]

    Visual instruction tuning,

    H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,” Advances in neural information processing systems, vol. 36, pp. 34 892– 34 916, 2023

    2023

  41. [41]

    Openai gpt-5 system card,

    A. Singh, A. Fry, A. Perelman, A. Tart, A. Ganesh, A. El-Kishky, A. McLaughlin, A. Low, A. Ostrow, A. Ananthramet al., “Openai gpt-5 system card,”arXiv preprint arXiv:2601.03267, 2025

    Pith/arXiv arXiv 2025

  42. [42]

    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,”CVPR, pp. 1685–1694, 2016

    2016

  43. [43]

    Image de-raining using a conditional generative adversarial network,

    H. Zhang, V . A. Sindagi, and V . M. Patel, “Image de-raining using a conditional generative adversarial network,”TCSVT, vol. 30, pp. 3943– 3956, 2017

    2017

  44. [44]

    Density-aware single image de-raining using a multi-stream dense network,

    H. Zhang and V . M. Patel, “Density-aware single image de-raining using a multi-stream dense network,”CVPR, pp. 695–704, 2018

    2018

  45. [45]

    Desnownet: Context-aware deep network for snow removal,

    Y .-F. Liu, D.-W. Jaw, S.-C. Huang, and J.-N. Hwang, “Desnownet: Context-aware deep network for snow removal,”IEEE Transactions on Image Processing, vol. 27, no. 6, pp. 3064–3073, 2018

    2018

  46. [46]

    Jstasr: Joint size and transparency-aware snow removal algorithm based on modified partial convolution and veiling effect removal,

    W.-T. Chen, H.-Y . Fang, J.-J. Ding, C.-C. Tsai, and S.-Y . Kuo, “Jstasr: Joint size and transparency-aware snow removal algorithm based on modified partial convolution and veiling effect removal,” inComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXI 16. Springer, 2020, pp. 754–770

    2020

  47. [47]

    All snow removed: Single image desnowing algo- rithm using hierarchical dual-tree complex wavelet representation and contradict channel loss,

    W.-T. Chen, H.-Y . Fang, C.-L. Hsieh, C.-C. Tsai, I. Chen, J.-J. Ding, S.-Y . Kuoet al., “All snow removed: Single image desnowing algo- rithm using hierarchical dual-tree complex wavelet representation and contradict channel loss,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 4196–4205

    2021

  48. [48]

    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,”TIP, vol. 28, no. 1, pp. 492–505, 2018

    2018

  49. [49]

    Deep multi-scale convolutional neural network for dynamic scene deblurring,

    S. Nah, T. H. Kim, and K. M. Lee, “Deep multi-scale convolutional neural network for dynamic scene deblurring,”CVPR, pp. 257–265, 2016

    2016

  50. [50]

    Human- aware motion deblurring,

    Z. Shen, W. Wang, X. Lu, J. Shen, H. Ling, T. Xu, and L. Shao, “Human- aware motion deblurring,”2019 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 5571–5580, 2019

    2019

  51. [51]

    A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics,

    D. Martin, C. Fowlkes, D. Tal, and J. Malik, “A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics,” inProceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, vol. 2, 2001, pp. 416–423 vol.2

    2001

  52. [52]

    Single image super-resolution from transformed self-exemplars,

    J.-B. Huang, A. Singh, and N. Ahuja, “Single image super-resolution from transformed self-exemplars,” in2015 IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR), 2015, pp. 5197–5206

    2015

  53. [53]

    Kodak lossless true color image suite,

    R. Franzen, “Kodak lossless true color image suite,”source: http://r0k. us/graphics/kodak, vol. 4, no. 2, p. 9, 1999

    1999

  54. [54]

    The unreasonable effectiveness of deep features as a perceptual metric,

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

    2018

  55. [55]

    Blind/referenceless image spatial quality evaluator,

    A. Mittal, A. K. Moorthy, and A. C. Bovik, “Blind/referenceless image spatial quality evaluator,” in2011 conference record of the forty fifth asilomar conference on signals, systems and computers (ASILOMAR). IEEE, 2011, pp. 723–727

    2011

  56. [56]

    Referenceless prediction of per- ceptual fog density and perceptual image defogging,

    L. K. Choi, J. You, and A. C. Bovik, “Referenceless prediction of per- ceptual fog density and perceptual image defogging,”IEEE Transactions on Image Processing, vol. 24, no. 11, pp. 3888–3901, 2015

    2015

  57. [57]

    Nima: Neural image assessment,

    H. Talebi and P. Milanfar, “Nima: Neural image assessment,”IEEE transactions on image processing, vol. 27, no. 8, pp. 3998–4011, 2018

    2018

  58. [58]

    Intra and inter parser- prompted transformers for effective image restoration,

    C. Wang, J. Pan, L. Wang, and W. Wang, “Intra and inter parser- prompted transformers for effective image restoration,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 7, 2025, pp. 7609–7618

    2025

  59. [59]

    Msp- former: Multi-scale projection transformer for single image desnowing,

    S. Chen, T. Ye, Y . Liu, T. Liao, J. Jiang, E. Chen, and P. Chen, “Msp- former: Multi-scale projection transformer for single image desnowing,” inICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

    2023

  60. [60]

    Global modeling matters: A fast, lightweight, and effective baseline for efficient image restoration,

    X. Jiang, N. Gao, H. Dou, X. Zhang, X. Zhong, Y . Deng, and H. Li, “Global modeling matters: A fast, lightweight, and effective baseline for efficient image restoration,”IEEE Transactions on Image Processing, vol. 35, pp. 2740–2754, 2026

    2026

  61. [61]

    Deep unfolding network for image desnowing with snow shape prior,

    X. Guo, X. Wang, X. Fu, and Z.-J. Zha, “Deep unfolding network for image desnowing with snow shape prior,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 35, no. 5, pp. 4740–4752, 2025

    2025

  62. [62]

    Dea-net: Single image dehazing based on detail-enhanced convolution and content-guided attention,

    Z. Chen, Z. He, and Z.-M. Lu, “Dea-net: Single image dehazing based on detail-enhanced convolution and content-guided attention,”IEEE Transactions on Image Processing, 2024

    2024

  63. [63]

    Emphasizing crucial features for efficient image restoration,

    H. Gao, B. Ma, Y . Zhang, J. Yang, J. Yang, X. Wang, and D. Dang, “Emphasizing crucial features for efficient image restoration,”Pattern Recognition, p. 113575, 2026

    2026

  64. [64]

    Omni-deblurring: Capturing omni-range context for image deblurring,

    Y . Li, H. An, T. Zhang, X. Chen, B. Jiang, and J. Pan, “Omni-deblurring: Capturing omni-range context for image deblurring,”IEEE Transactions on Circuits and Systems for Video Technology, pp. 1–1, 2025

    2025

  65. [65]

    A polarization- aided transformer for image deblurring via motion vector decomposi- tion,

    D. Chen, S. Zhou, J. Pan, J. Shi, L. Qu, and J. Yang, “A polarization- aided transformer for image deblurring via motion vector decomposi- tion,” inProceedings of the Computer Vision and Pattern Recognition Conference (CVPR), June 2025, pp. 28 061–28 070

    2025

  66. [66]

    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,” inCVPR, 2022

    2022

  67. [67]

    All-in-one transformer for image restoration under adverse weather degradations,

    J. Mao, Y . Yang, X. Yin, L. Shao, and H. Tang, “All-in-one transformer for image restoration under adverse weather degradations,”IEEE Trans- actions on Pattern Analysis and Machine Intelligence, vol. 48, no. 6, pp. 6628–6641, 2026

    2026

  68. [68]

    Toward real-world single image deraining: A new benchmark and beyond,

    W. Li, Q. Zhang, J. Zhang, Z. Huang, X. Tian, and D. Tao, “Toward real-world single image deraining: A new benchmark and beyond,”arXiv preprint arXiv:2206.05514, 2022

    Pith/arXiv arXiv 2022

  69. [69]

    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 transactions on image processing, vol. 28, no. 1, pp. 492–505, 2018

    2018

  70. [70]

    A high-quality denoising dataset for smartphone cameras,

    A. Abdelhamed, S. Lin, and M. S. Brown, “A high-quality denoising dataset for smartphone cameras,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 1692–1700

    2018

  71. [71]

    Dswinir: Rethinking window-based attention for image restoration,

    G. Wu, J. Jiang, K. Jiang, X. Liu, and L. Nie, “Dswinir: Rethinking window-based attention for image restoration,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 48, no. 4, pp. 4350– 4366, 2026

    2026

  72. [72]

    Reference-based multi-stage progressive restoration for multi-degraded images,

    Y . Zhang, Q. Yang, D. M. Chandler, and X. Mou, “Reference-based multi-stage progressive restoration for multi-degraded images,”IEEE Transactions on Image Processing, vol. 33, pp. 4982–4997, 2024

    2024