acceptodds
Under review as a conference paper at ICLR 2027

Fine-Grained Degradation Representations for Efficient Agentic Image Restoration

Abstract

Existing agentic image restoration methods rely on coarse degradation labels, which cannot capture fine-grained characteristics within the same degradation type and may lead to inefficient tool selection. We propose Locality-prior Agentic Image Restoration (LAIR), an efficient framework that reuses successful restoration trajectories based on fine-grained degradation similarity. LAIR employs a Degradation-Query Momentum Contrastive Encoder (DQ-MCE), which combines learnable degradation queries, cross-content momentum contrastive learning, and parameter-aware regularization to learn content-robust yet degradation-sensitive representations. These representations enable the agent to retrieve, rank, execute, and verify relevant historical tool chains instead of repeatedly rolling back and rescheduling when restoration fails. Experiments on both synthetic and real-world degradations show that LAIR improves restoration fidelity while maintaining competitive perceptual quality. It also reduces runtime by up to 58% and requires fewer tool invocations than AgenticIR. These results demonstrate that fine-grained degradation representations provide an effective bridge between historical restoration trajectories and efficient agentic decision-making.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Restoration, Agent, Low-level Vision

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_7EpgOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LAIR is an agentic image restoration framework that retrieves and reuses historical restoration trajectories using fine-grained degradation representations. Its DQ-MCE encoder combines learnable degradation queries, cross-content momentum contrastive learning, and synthetic degradation-parameter regression. Retrieved trajectories are filtered by degradation compatibility, ranked by stored quality, and verified after execution. The system falls back to AgenticIR planning when reuse fails and updates its memory with successful results. Experiments report modest fidelity improvements and substantial online runtime reductions on synthetic and real-world restoration data.

Critical strengths
  1. The motivation is practical and well supported. Coarse degradation labels can group images that require different restoration tools. Learning finer representations for experience retrieval directly addresses this limitation.

  2. LAIR reduces runtime from 156.91 to 65.90 seconds per image in the reported synthetic comparison. Its integration into 4KAgent and evaluation on real-world images provide additional evidence that memory reuse can reduce execution cost.

Critical weaknesses
  1. The novelty over existing memory-based agents is incremental. SEAR already retrieves restoration trajectories using degradation-aware fingerprints. A matched comparison using the same memory, verification rules, and budget would better isolate the benefit of DQ-MCE.

  2. The representation ablations are too narrow. Retrieval ablations cover only 100 rain-plus-dark images. Separate tests of degradation queries, cross-content pairing, momentum learning, and parameter-only training are needed to establish which components matter.

  3. The online-memory evaluation protocol needs clarification. Updating memory during testing makes results potentially dependent on image order and previous queries. Shuffled-order repeats, memory-reset settings, and performance versus memory size would improve reproducibility.

  4. The role of offline supervision and preparation cost is insufficiently isolated. Initial memories use reference-based selection of the best tool chains, while the encoder requires substantial synthetic training. Comparisons should match memory supervision and distinguish online savings from offline preparation costs.

Overall, I lean toward acceptance because the paper presents a coherent representation-learning approach with clear online efficiency benefits and useful supporting diagnostics. The contribution is incremental, but practically relevant. Stronger component ablations and a reproducible online-memory protocol would substantially strengthen the evidence.

Questions for authors

See weaknesses.

5 comments

Thank you for recognizing LAIR’s efficiency gains and the evidence from the retrieval ablations.

W1. The comparison with SEAR

For this response, we independently implemented SEAR because we could not locate an official public implementation. We evaluated the independent implementation at the reported P-MCTS budget multiplier (=9[official parameter in paper]) and at (=7). The Group B comparison is shown below because it is also the split used for SEAR's published efficiency analysis, while Severity-L2 represents the variant that replaces only LAIR’s retrieval representation and matching with SEAR’s fingerprint-based retrieval, while keeping the initial memory, restoration tools, verification rules, planning budget, and memory-update policy unchanged.

The tool call per image in official SEAR paper in Group B is 8.15 instead of 9.34 or 9.74 in our implementation. However, I guess the run time will also longer than LAIR. LAIR shows best fidelity and better efficiency.

Group B (400 images)PSNR-Y ↑SSIM-Y ↑LPIPS ↓MANIQA ↑CLIP-IQA ↑MUSIQ ↑Tool invocation ↓Time(s)/image ↓
LAIR22.4498\textbf{22.4498}0.7325\textbf{0.7325}0.2939‾\underline{0.2939}0.34270.34270.49810.498161.0305\textbf{61.0305}4.08\textbf{4.08}62.61\textbf{62.61}
SEAR (ω\omega=7)21.256521.25650.72480.72480.29730.29730.33540.33540.5195‾\underline{0.5195}60.9647‾\underline{60.9647}9.349.34176.20176.20
SEAR (ω\omega=9)21.281221.28120.72460.72460.29510.29510.3453‾\underline{0.3453}0.5267\textbf{0.5267}60.197260.19729.749.74176.75176.75
SEAR(official paper)22.1332‾\underline{22.1332}0.7251‾\underline{0.7251}0.2890\textbf{0.2890}0.3447{0.3447}0.51340.513460.652060.65208.158.15−-
Severity-L221.547221.54720.72360.72360.29430.29430.3475\textbf{0.3475}0.49300.493060.243360.24336.29506.295064.3464.34

W2. more ablation on DQ-MCE

We agree that the original 100-image rain+dark experiment provided limited coverage. We now report separate component ablations covering all four compound-degradation settings in Group B (400 images). For each split, 80 images per setting form the memory bank (320 total), and the remaining 20 per setting serve as restoration queries (80 total). We repeat evaluation over ten memory/query splits using seeds 0–9, keeping each variant’s trained checkpoint fixed and disabling planning fallback to assess memory reuse. The new tables report averages across these splits.

We evaluate five ablations: (i) query-based aggregation: replacing learnable queries and cross-attention with mean-pooled patch features and seven independent projection MLPs, approximately matching parameter count while preserving the 7×5127\times512 output; (ii–iii) training objectives: using InfoNCE only or parameter-regression MSE only; (iv) cross-content pairing: selecting positives from another degraded version of the same clean crop, while retaining ordered-degradation matching and nearest-parameter pairing; and (v) EMA updates: setting m=0m=0 while retaining stop-gradient keys and the negative queue; We also include Severity-L2, mean-patch, and CLS-token feature baselines.

ExtractorPSNR ↑SSIM ↑Recall@5 (%) ↑Rank ↓
DQ-MCE24.56\mathbf{24.56}0.7742\mathbf{0.7742}58.50\mathbf{58.50}5.70\mathbf{5.70}
w/o queries24.44‾\underline{24.44}0.7717‾\underline{0.7717}50.1350.136.13‾\underline{6.13}
Parameter only24.4024.400.76960.769649.5049.506.646.64
Contrastive only23.9723.970.76560.765655.5055.506.866.86
w/o cross-content23.8123.810.75820.758238.1338.136.926.92
w/o EMA23.1923.190.73040.730425.7525.7510.2710.27
Severity-L224.1224.120.76920.769257.88‾\underline{57.88}6.426.42
Mean patch23.7823.780.75580.755833.0033.007.147.14
CLS token23.6923.690.75660.756632.1332.137.317.31

The table shows the training objectives are necessary to help DQ-MCE extract fine-grained features, achiving better performance than discrete-labels.

W3. online-memory evaluation protocol

Our initial memory is constructed from 160 Group A exploration samples following the split of AgenticIR. These samples are excluded from the 1,440-image test set. Before each independent evaluation run, we restore exactly the same initial memory; online updates are retained within that run but are not carried over to subsequent runs.

The original results use a fixed ascending image order(from id 1 to id 100). To assess order sensitivity, we additionally evaluated all 1,440 test images under ten shuffled orders (seeds 0–9), changing only the test-image order.

seedPSNR-Y ↑SSIM-Y ↑LPIPS ↓MANIQA ↑CLIP-IQA ↑MUSIQ ↑
021.96930.68970.33320.32400.481558.0830
121.94650.68810.33360.32580.483758.3614
221.93540.68860.33410.32600.481258.5655
322.03710.68940.33500.32580.481358.1931
421.92410.68790.33450.32340.481358.2458
521.91390.68740.33500.32510.481958.3294
621.85860.68710.33770.32260.481158.0811
722.04190.69100.33130.32610.480858.3314
821.98280.68850.33470.32540.481758.1955
921.99570.68920.33450.32630.481458.2465
mean ±std21.9605 ± 0.05680.6887 ± 0.00120.3344 ± 0.00160.3251 ± 0.00130.4816 ± 0.000858.2633 ± 0.1437
Δ(main experiments − mean)+0.0611−0.0002−0.0035+0.0015−0.0006+0.2446

These results show small aggregate metric variation across the tested permutations, which demonstrates our LAIR's robustness in shuffled-order repeats.

Regarding performance versus memory size, we have added a comparison of initial memory sizes (0, 80, and 160 entries) to the supplementary material (page 18 A.4), with online updates enabled in all settings. Memory initialization generally improves early-stage PSNR-Y compared with starting from an empty memory. The mean PSNR-Y gap between the 160- and 0-entry initializations decreases from 0.624 dB over images 1–40 to 0.240 dB over images 41–80. These results suggest reduced sensitivity to initialization during later processing because of the online memory growth, although the gaps fluctuate and some differences remain. The table below reports PSNR-Y (dB), averaged over successive, non-overlapping blocks of 10 images per degradation combination and then equally across the eight combinations.

step080160
1–1022.589923.270123.2505
11–2021.453021.530321.6369
21–3021.273822.212022.3338
31–4021.135421.095021.7263
41–5023.787523.836824.0892
51–6022.722922.787322.8520
61–7022.485822.688922.5473
71–8020.647921.160321.1156

W4. Clarification of preparation cost

LAIR and AgenticIR share the offline cost of exhaustive tool-chain execution on the same 160 exploration images to get inital memory(or static experience).The reported runtime reduction concerns online inference latency; it excludes offline preparation and encoder training.

However, for training cost, DQ-MCE requires an additional 2–3 hours of training on eight RTX A6000 GPUs, performed only once. But users benefit from roughly halved average waiting times and a more responsive restoration experience.

Reviewer_JEQXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes E²IR-Agent, an experience-driven self-evolving framework for image restoration agents, aiming to improve restoration quality and inference efficiency in scenarios involving multiple coupled degradations. The authors formulate restoration experience along three dimensions: tool selection, degradation removal order, and visual quality preference. Based on this formulation, they introduce a hierarchical experience pool that organizes restoration strategies at different granularities, enabling degradation-pattern-aware planning. Furthermore, a self-evolving mechanism is developed to continuously update the experience pool through quality-based ranking and pattern profile learning. Experiments on synthetic and real-world datasets demonstrate improvements in restoration quality and reduced tool invocations compared with existing image restoration agents.

Critical strengths
  1. Improved efficiency. LAIR substantially reduces runtime and tool invocations while maintaining competitive restoration quality.

  2. Improved effectiveness. Fidelity gains and retrieval ablations support DQ-MCE’s ability to match degraded inputs with suitable restoration trajectories.

Critical weaknesses
  1. Limited novelty relative to prior work. UtilityIR already models degradation type and severity with continuously adjustable features, while SEAR retrieves and reuses tool chains based on degradation states. The additional novelty of DQ-MCE beyond these established ideas therefore requires stronger justification.

  2. Limited scope of contribution. The contribution primarily consists of designing and training DQ-MCE and using its features to guide tool-chain selection within an existing agentic restoration pipeline. This relatively narrow methodological scope limits the significance of the overall contribution.

  3. Insufficient evidence of advantages over SOTA methods. The paper lacks a direct efficiency comparison with SEAR, while restoration quality improvements are limited to selected metrics. Consequently, a superior quality–efficiency trade-off over this closely related method remains insufficiently demonstrated.

Questions for authors
  1. Could you add a flowchart in Section 3.4 illustrating the complete agentic restoration loop, including retrieval, verification, fallback planning, and memory updates?

  2. Could you compare LAIR with SEAR under identical settings in terms of both efficiency and restoration quality? Please also include no-reference quality metrics in Tables 2 and 4.

  3. Since the new contributions primarily concern degradation representation and trajectory retrieval, could you clarify their relationship with the existing VLM/LLM components and their contribution to agentic decision-making?

  4. Why learn degradation representations for retrieving historical trajectories from AgenticIR rather than directly learn a tool-chain selection policy from the available training data?

  5. Could you extend the 4KAgent integration experiment in Table 2 with additional restoration-quality metrics to establish whether LAIR’s quality improvements generalize beyond the AgenticIR backbone?

3 comments

Thank you for recognizing LAIR’s efficiency gains and the evidence from the retrieval ablations.

Initially I would clarify that we began developing LAIR before SEAR first appeared on arXiv on June 27, 2026, which could be proven by git log when code is released.

W1. Novelty relative to UtilityIR and SEAR. Our contribution includes an explicit investigation of how degradation representation affects the retrieval and reuse of restoration tool chains. Specifically, we ask whether fine-grained learned degradation features(by our DQ-MCE) provide useful information beyond coarse type-and-severity descriptors for transferring restoration trajectories across images with different content. DQ-MCE addresses this question through degradation-specific representations learned from cross-content image pairs and known synthesis parameters. We evaluate their utility through two complementary experiments: a fixed-memory comparison with the same top-5 retrieval budget, and an online ablation in which both variants start from the same memory and subsequently update it independently. In the online evaluation on 1,440 images, replacing Severity-L2 with DQ-MCE improves PSNR-Y by 1.05 dB and reduces tool attempts after prefix deduplication by 4.9%, supporting the practical value of the learned representation within the restoration workflow.

LAIR retrieval methodPSNR-Y ↑SSIM-Y ↑LPIPS ↓MANIQA ↑CLIP-IQA ↑MUSIQ ↑Tool invocation ↓Time(s)/image ↓
Severity-L220.974120.97410.67560.67560.33530.33530.3280\textbf{0.3280}0.47030.470356.678656.67865.11395.113984.7384.73
DQ-MCE22.0216\textbf{22.0216}0.6885\textbf{0.6885}0.3309\textbf{0.3309}0.32650.32650.4810\textbf{0.4810}58.5079\textbf{58.5079}4.8611\textbf{4.8611}80.22\textbf{80.22}

W2. Scope of the contribution. LAIR can also adds a verified trajectory-reuse branch to an existing agent. We demonstrate this integration in native 4KAgent. Mean tool attempts fall by 49.6% and mean workflow time by 24.6%. The same interface can be used with a trained agent that exposes tool execution, verification, and fallback.

And the learned features also support degradation perception across image contents which can provide fine-grained features for futher restoration work.

W3 & Q2. Quality–efficiency comparison with SEAR.

However, for this response, we independently implemented SEAR because we could not locate an official public implementation. We evaluated the independent implementation at the reported P-MCTS budget multiplier $$\omega=9[officialparameterinpaper])andat(=9[official parameter in paper]) and at (\omega=7=7. The Group B comparison is shown below because it is also the split used for SEAR's published efficiency analysis.

The tool call per image in official SEAR paper in Group B is 8.15 instead of 9.34 or 9.74 in our implementation. However, I guess the run time will also longer than LAIR.

Group B (400 images)PSNR-Y ↑SSIM-Y ↑LPIPS ↓MANIQA ↑CLIP-IQA ↑MUSIQ ↑Tool invocation ↓Time(s)/image ↓
LAIR22.4498\textbf{22.4498}0.7325\textbf{0.7325}0.2939\textbf{0.2939}0.3427‾\underline{0.3427}0.49810.498161.0305\textbf{61.0305}4.08\textbf{4.08}62.61\textbf{62.61}
SEAR (ω\omega=7)21.256521.25650.7248‾\underline{0.7248}0.29730.29730.33540.33540.5195‾\underline{0.5195}60.9647‾\underline{60.9647}9.349.34176.20176.20
SEAR (ω\omega=9)21.2812‾\underline{21.2812}0.72460.72460.29510.29510.3453\textbf{0.3453}0.5267\textbf{0.5267}60.197260.19729.749.74176.75176.75

Q1. Flowchart for the complete restoration loop. We add a flowchart to §3.4(page 6 in the latest version) showing how DepictQA identifies the active degradations, how DQ-MCE retrieves candidate trajectories, and how LAIR filters and orders those candidates before execution. The figure will show the verification rule explicitly. LAIR accepts a candidate immediately at residual severity grade \Gamma$=5$. It retains candidates at \Gamma=4=4 and selects among them if none reaches $$\Gamma=5=5. If no candidate qualifies, LAIR invokes the original planning workflow. Both branches lead to a final output and a validation step for writing eligible trajectories back to memory.

Q3. Relationship to the VLM and LLM. DepictQA identifies the active degradations and verifies the restored output. DQ-MCE determines which historical trajectories enter the candidate set, and their stored quality grades determine the order in which they are tried. The verifier then determines whether a candidate can supply the final output. When no retrieved chain qualifies, the existing LLM planner orders the restoration tasks and try tools to restore images(fallback). The 45.97% direct reuse rate on synthetic images quantifies how often retrieval supplies the executed chain and bypasses fallback planning.

Q4. Why retrieve trajectories instead of learning a policy directly? Because the cost to sample trajectory is expensive. Historical enumeration logs average 579 s for two-degradation images and 6,437 s for three-degradation images in the earlier exploration. These measurements characterize the cost of acquiring additional trajectories. Besides, the training cost of VLM is also expensive, TIR-Agent, for example, uses 39,827 training samples containing 4KAgent trajectories for supervised fine-tuning. It fully fine-tunes Qwen3-VL-8B on eight A100s and then trains the policy with RL on eight A100s plus 32 RTX 3090s that run restoration tools during rollouts.

Howevery, choosing appropriate successful trajectories by retrieving could improve performance without too much cost. By the way, our method can also used in training-based agents because it's a plug-in module.

Q5. Additional quality metrics for 4KAgent.

We extended the comparison in Table 2 (p. 8, line 405) to report six restoration-quality metrics. Adding memory-guided trajectory reuse to native 4KAgent reduces mean tool invocations from 11.22 to 5.65 (49.6%) and mean workflow time from 256.15 to 193.19 seconds per image (24.6%). These results demonstrate that the efficiency benefits extend beyond AgenticIR, with a quality trade-off on 4KAgent. This is because there is complex quality verification in 4KAgent to ensure the performance so that the memory reuse in 4KAgent primarily reduce the runtime.

MethodPSNR-Y ↑SSIM-Y ↑LPIPS ↓MANIQA ↑CLIP-IQA ↑MUSIQ ↑Tool invocations/image ↓Time (s/image) ↓
4KAgent21.00780.67670.30090.37210.556162.691811.2206256.15
4KAgent with memory reuse21.00810.66160.31100.37160.545662.29095.6516193.19

We have added the flowchart and additional quality metrics for 4KAgent in the paper. Let us thank you again for your thoughtful and constructive review. With only a short time remaining before the discussion window closes, we wanted to gently check in to ask whether our clarifications have resolved your concerns, or if there are any remaining questions you would like to discuss. If anything is still unclear, please feel free to let us know when you have a moment, as we would truly welcome the opportunity to address your doubts as thoroughly as possible in the remaining hours.

Thank you very much for your time and consideration.

Reviewer_2LfoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LAIR, which reuses historical restoration tool chains using fine-grained degradation representations. Its core component, DQ-MCE, combines degradation queries, cross-content contrastive learning, and parameter regression to capture fine-grained degradation characteristics. Experiments on synthetic and real-world datasets demonstrate competitive restoration quality with fewer tool invocations and reduced runtime.

Critical strengths

The motivation is clear, as images with the same predicted degradation severity may favor different optimal restoration tools.

Critical weaknesses
  1. Figures 4–5 provide evidence of tool-chain locality, and Appendix A.2 reports memory-reuse acceptance rates. However, these results do not directly quantify the performance gap between retrieved chains and the best chain for each image in the evaluated search space. The nearest-neighbor optimal tool-chain agreement rate is 28.14%, above the random baseline, but label agreement alone does not establish transfer quality. Could the authors evaluate transfer regret relative to this per-image optimum and analyze how transfer success and regret vary with feature distance?

  2. Line352 describes AgenticIR, 4KAgent, and SEAR as "training-free agent methods," whereas LAIR requires additional DQ-MCE training with LORA and learnable degradation queries. The comparison is unfair without accounting for this additional training cost. The authors should also compare LAIR with training-based agentic methods.

Questions for authors
  1. Cross-content momentum contrastive learning aims to suppress content-dependent information in degradation representations. However, Figure 3 mainly demonstrates sensitivity to degradation parameters rather than content invariance. Could the authors evaluate how much image content information remains in the frozen degradation representations?

  2. Given recent advances in general-purpose image editing models such as GPT-Image-2.5, could the authors clarify the advantages of specialized agentic restoration over end-to-end editing? A brief discussion or a small-scale experiment would help clarify the value of the agentic approach.

6 comments

Thank you for raising these questions. We report new results on tool-chain transfer, cross-content degradation retrieval, and direct generative editing. We also clarify the offline supervision and training costs of LAIR.

W1. Tool-chain transfer regret. We recomputed nearest neighbors across all the images using the original Fig. 5 grouping rules. For each query and metric, we measured how far the transferred chain’s result was from the best value achieved by any evaluated chain for that query. Nearest-neighbor transfer had lower mean regret than random selection from feasible memory entries on all six metrics.

MetricNearest-neighbor transfer ↓Random feasible memory ↓
PSNR-Y (dB)1.7742.678
SSIM-Y0.04970.0667
LPIPS0.06380.0807
MUSIQ1.3591.532
CLIP-IQA0.07200.0812
PSNR-RGB (dB)1.6372.494

W2. Offline training cost and integration with existing agents. LAIR trains DQ-MCE once as a reusable retrieval encoder. Its training supervision comes from applying known degradations to clean images and forming cross-content pairs. We freeze the pretrained DINOv3 weights and train rank-8 LoRA adapters and the encoder heads for 12,000 steps on eight RTX A6000 GPUs.

Training a tool-calling policy entails substantially more work. TIR-Agent, for example, uses 39,827 training samples containing 4KAgent trajectories for supervised fine-tuning. It fully fine-tunes Qwen3-VL-8B on eight A100s and then trains the policy with RL on eight A100s plus 32 RTX 3090s that run restoration tools during rollouts. Our timing estimates illustrate the cost of collecting demonstrations alone. Based on measured tool runtimes, obtaining the final AgenticIR trajectory took an estimated 159 s per image for two degradations and 356 s for three, including search and backtracking. DQ-MCE learns from synthetic degradations without executing these restoration trajectories during encoder training which only use eight A6000s for about 2 hours. Populating the retrieval memory is a separate offline cost, and Table 2 reports online restoration time.

LAIR is also a plug-in for an existing restoration agent. In Table 2, we insert the same retrieval module between 4KAgent’s proposal and tool execution stages. Retrieved chains run through 4KAgent’s own tools and verifier, with control returning to the original agent when reuse fails. This reduced mean tool calls from 11.22 to 5.70 and runtime from 256.15 to 193.19 s per image . PSNR was 21.01 dB in both settings, and SSIM decreased from 0.6767 to 0.6616. The same interface can be used with an already trained agent by retaining its learned policy as the fallback, without retraining that policy.

Thank you for your response and additional experiments. Here are my remaining concerns.

W2:

First, the authors only compare LAIR with training-free agentic methods, although LAIR itself is not training-free. I believe the authors should also compare LAIR's performance with training-based agentic methods, such as TIR-Agent mentioned in the rebuttal, rather than only comparing training costs. I think LAIR's positioning remains unclear without such comparisons.

Second, the authors claim that memory reuse improves the efficiency of 4KAgent. However, Table 2 shows that this comes at the cost of degraded performance on several metrics. This suggests that memory reuse does not always lead to improvements.

Q1:

The additional experiment supports content-robust degradation retrieval. However, as the authors acknowledge, "The amount of source-image information still decodable from the frozen representation remains open." Thus, it remains unclear how much image content information is retained in the learned representations, which was the main point of my original question.

W2. Initially, thank you for raising this concern. We agree that performance comparisons with training-based agents would help clarify LAIR's positioning.

LAIR and TIR-Agent differ in what is learned, not merely in training cost. TIR-Agent trains the agent's tool-calling policy through SFT and RL. LAIR trains DQ-MCE for degradation-based trajectory retrieval, retaining the original planning policy as the fallback. Training a retrieval encoder does not make LAIR a policy-training method. Comparisons of existing agents with and without LAIR directly test our contribution.

LAIR could also complement trained policies; However, TIR-Agent's implementation was not publicly available for controlled integration experiments.

We provide the following comparison of reported Group A results. Group A contains degradation combinations represented in LAIR's initial memory: following AgenticIR's predefined split, 160 exploration images initialize the trajectory memory, while the remaining 640 images are held out for evaluation . These exploration images are not used to train the VLM.

MethodPSNR ↑SSIM ↑LPIPS ↓MANIQA ↑CLIP-IQA ↑MUSIQ ↑
LAIR22.44400.71570.30160.33350.482660.3328
TIR-Agent22.070.69350.28740.39070.545463.69

TIR-Agent values are quoted from its paper.

LAIR has higher reported PSNR and SSIM, while TIR-Agent has stronger LPIPS and no-reference scores. This comparison positions LAIR as a memory-reuse approach that achieves competitive restoration quality without additional VLM policy training.

Second, Table 2 supports the claimed efficiency improvement. Relative to the original 4KAgent:

MetricChange
PSNR ↑Effectively unchanged:21.01 dB
SSIM ↑−2.23%
LPIPS ↓+3.36%
MANIQA ↑−0.13%
CLIP-IQA ↑−1.89%
MUSIQ ↑−0.64%
Tool invocations ↓−49.63%
Runtime ↓−24.58%

PSNR remains effectively unchanged, while all other quality-metric changes are below 3.4% , alongside nearly halved tool calls and approximately one-quarter lower runtime(efficient). These modest metric changes do not negate the substantial efficiency gain. Our claim concerns reduced execution cost, not simultaneous improvement across every quality metric.

Q1. We thank the reviewer for this suggestion. We agree that Figure 3 mainly demonstrates degradation-parameter sensitivity. To assess residual semantic information, we futher conduct object- and scene-classification probes using clean images. Both encoders remain frozen; only a two-layer MLP with a 256-unit hidden layer is trained.

Datasettoken + MLPOriginal DINO + MLP
Caltech-10125.30 ± 0.23%97.75 ± 0.43%
MIT Indoor-6722.13 ± 0.14%90.93 ± 0.25%

Within each run, we average the mean per-class Top-1 accuracies across the seven degradation tokens. We report the mean ± standard deviation of these averages over three runs. The substantially lower accuracies indicate reduced semantic decodability.

We additionally repeat the cross-content similarity analysis using more LSDIR source images(2,000). Active-token cosine similarity is 0.9297 ± 0.0665 for different-content, same-degradation pairs, versus 0.1447 ± 0.0946 for same-content, different-degradation distractors, providing complementary evidence of content robustness.

These findings support reduced content dependence , consistent with our goal of suppressing content-dependent variations.

Comments on the paper

5 comments

Q1. Cross-content degradation retrieval. We evaluated frozen DQ-MCE on 4,500 queries generated from 250 clean 512×512 LSDIR validation images, with six degradation types at three strengths per image. We concatenated the seven normalized degradation tokens and L2-normalized the resulting descriptor without further training. Each gallery contained a positive from a different source image with the same degradation type and parameters, a distractor from the same clean source image with a different degradation type, and 15–16 additional negatives that differed in both source image and degradation type. We excluded candidates of the same type at other strengths to isolate discrimination between degradation types.

The mean cosine similarities were 0.9349 for positives and 0.1504 for same-content distractors. The positive ranked above the same-content distractor in 99.91% of queries, and gallery Top-1 accuracy was 99.67% . Thus, matching degradation dominated the tested retrieval rankings. The amount of source-image information still decodable from the frozen representation remains open.

Q2. Fidelity versus direct generative editing. We compared LAIR with GPT Image 2.5 Sunburst on 16 randomly sampled test images, covering all 16 degradation settings, using both generic restoration prompts and prompts specifying the degradation types.

MethodPSNR-Y ↑SSIM-Y ↑LPIPS ↓CLIP-IQA ↑
LAIR22.800.72140.30900.5059
GPT, generic prompt20.240.63040.27220.5657
GPT, known degradation types19.870.61410.26120.5763

GPT’s strong generative prior can produce visually plausible details and textures that are inconsistent with the original image. Supplying the degradation types still leads to this content alteration. LAIR more faithfully preserves the original display content and surrounding structures while removing the degradation.

Sorry, the run time reduced by LAIR module in 4KAgent is 193.19s per image. I made a mistake here.

Thank you for your recognizing LAIR's contribution!

W1&Q1. structure limitation.

We thank the reviewer for highlighting this limitation. LAIR could not recognize degradation beyond the vocabulary or go into wrong perception. Following the predefined degradation setting of prior agentic restoration work such as AgenticIR, our implementation uses a fixed DepictQA vocabulary and corresponding degradation queries in DQ-MCE. Therefore, LAIR cannot explicitly diagnose a previously unseen degradation category or represent it with a dedicated query. Fallback planning remains restricted to recognized tasks and available tools, and cannot compensate for this limitation. We agree that this scope should have been explicitly acknowledged and will clarify it in the revised manuscript.

Supporting the degradation types mentioned by the reviewer would require extending DQ-MCE with corresponding degradation queries and training data, fine-tuning DepictQA for their diagnosis and restoration assessment, and incorporating suitable restoration tools and experiences. DQ-MCE allows parameter-efficient adaptation through LoRA adapters and query-related heads while keeping the pretrained backbone frozen, avoiding full backbone retraining. Memory embeddings would also be refreshed after encoder adaptation. And the scalability of our LAIR have been shown.

W3 & Q3 reproducibility

Yes, we will release code, model weights, synthetic degradation data, and memory banks. Now we have uploaded .zip codes in supplementary material. If the paper is accepted, we will release other parts.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.