acceptodds
Under review as a conference paper at ICLR 2027

ReVRA: Reflective Video Restoration Agent with Self-Corrective Dynamic Pipelines

Abstract

Degradations in real-world user-generated videos are inherently complex and multifaceted, making monolithic restoration tools struggle to decouple and resolve multiple, overlapping artifacts. While recent agentic restoration systems offer greater flexibility by chaining specialized restoration tools into pipelines, they suffer from three limitations: planning relies solely on the input video and static tool priors, making it difficult to select appropriate toolboxes; execution follows a fixed pipeline that cannot adapt to unexpected artifacts introduced by tools; and evaluation relies on absolute quality scores that poorly align with human perception. We propose ReVRA, a novel Reflective Video Restoration Agent that constructs adaptive restoration tool sequences through dynamic self-correction. ReVRA decomposes complex restoration into self-corrective stages that re-examine the current video state to adapt tool usage to unexpected artifacts. To improve planning, ReVRA replaces one-shot prediction with toolbox decisions guided by instance-aware evolving memory and verified by execution outcomes, which also refine the memory to provide increasingly accurate priors. Moreover, ReVRA evaluates restoration results via relative comparisons, better aligning with the human preference selection process. Experiments across various benchmarks spanning synthetic and real-world degradation scenarios demonstrate that ReVRA achieves superior restoration quality across a wide spectrum of complex degradations, validating the effectiveness of the reflective agentic framework for real-world video restoration.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
video restoration, restoration agent

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_3BEJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ReVRA, a video restoration agent that dynamically selects restoration tools through iterative planning, execution, and evaluation. It introduces an evolving memory for tool selection and a DPO-trained VLM for pairwise quality assessment. Experiments across seven benchmarks report improvements over existing restoration methods and agents.

Critical strengths
  1. The iterative plan–execute–evaluate framework is reasonable for handling compound degradations. Table 4 provides evidence that re-planning can improve restoration quality compared with fixed pipelines.

  2. The evaluation covers seven benchmarks, including real-world UGC videos. I appreciate the aligned-tool comparisons in Table 3 and the efficiency analysis of evolving memory in Table 6.

Critical weaknesses
  1. I noticed that the TaoUGC results for ReVRA in Table 1 exactly match the memory-free configuration in Table 6, rather than the proposed full IEM configuration. This raises questions about which system was actually evaluated. Also, the reported human preferences for brute-force and ReVRA (53%, 50%, and 3% ties) do not sum to 100%. I think these inconsistencies need clarification.

  2. Appendix A.6 averages CLIPIQA, MUSIQ, DOVER, and negative NIQE without normalization. Since MUSIQ has a much larger numerical range, this baseline is effectively dominated by MUSIQ. Moreover, Table 5 compares the DPO-trained evaluator against untuned VLMs, making it difficult to distinguish the benefit of pairwise comparison from preference fine-tuning. Interestingly, the untuned VLMs perform substantially worse than the metric-based baseline on KVQ.

Questions for authors
  1. Why do the TaoUGC results in Table 1 match the memory-free configuration in Table 6 rather than the full ReVRA configuration? Which results correspond to the final system?

  2. Why are the IQA/VQA metrics averaged without normalization in Appendix A.6? Could the authors clarify whether normalized or calibrated metric-based comparisons were evaluated?

  3. Could the authors provide an ablation separating pairwise comparison from DPO fine-tuning? I think this is important for validating the claimed advantage of the comparison-based evaluator.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 3BEJ (Part 1/3)

We thank the reviewer for the careful and detailed review, as well as for the positive recognition of our iterative plan–execute–evaluate framework and the breadth of our evaluation. Below we address the reviewer's concerns point by point: confirmation of the TaoUGC results in Table 1, the computation of the human preference statistics, an ablation separating the benefits of relative comparison from those of DPO training in the evaluation VLM, and a comparison of selection accuracy against normalized QA-based methods.


W1&Q1: Verification of TaoUGC Results in Table 1

I noticed that the TaoUGC results for ReVRA in Table 1 exactly match the memory-free configuration in Table 6, rather than the proposed full IEM configuration. This raises questions about which system was actually evaluated. Also, the reported human preferences for brute-force and ReVRA (53%, 50%, and 3% ties) do not sum to 100%. I think these inconsistencies need clarification.

Why do the TaoUGC results in Table 1 match the memory-free configuration in Table 6 rather than the full ReVRA configuration? Which results correspond to the final system?

Response: Table 1 evaluates the final system with IEM enabled. The TaoUGC row mistakenly copied the memory-free configuration results from Table 6, and we have corrected it to report the corresponding final-system results.

We thank the reviewer for identifying this error. The final system in Table 1 enables IEM, initializes it empty before evaluating each dataset, and updates it online during inference. However, the original TaoUGC row contained a transcription error, which we have corrected. To make clear which results correspond to which configuration, we list them below:

MetricTable 1: originally reported memory-free resultsTable 1: corrected empty-initialized online IEMTable 6: pre-initialized online IEM
NIQE ↓5.1355.1915.055
CLIPIQA ↑0.4900.4930.500
MUSIQ ↑63.7664.0664.86
DOVER ↑0.8180.8210.826

The corrected Table 1 results use an empty IEM initialization to prevent experience from carrying over across datasets. The full IEM configuration in Table 6 is instead pre-initialized on KVQ, GoPro, and VideoLQ before continuing online updates on TaoUGC. Both configurations use the same system components, but their initialization differs. This explains the difference between the corrected Table 1 results and the full IEM configuration in Table 6.

The corrected comparison on TaoUGC is shown below. CLIPIQA, MUSIQ, and DOVER remain higher than those of all compared methods; NIQE becomes slightly worse, but its ranking is unchanged. The correction therefore preserves the conclusions.

MetricsBasicVSR++MoCE-IRAverNetAgenticIR4KAgentOurs
NIQE↓4.8475.3585.5175.4365.3735.191
CLIPIQA↑0.4850.3760.3940.3740.3940.493
MUSIQ↑61.5155.6654.3555.7657.0964.06
DOVER↑0.8180.7530.7530.7470.7580.821

Revision: We have corrected the TaoUGC row in Table 1 and clarified the initialization settings in the captions of Tables 1 and 6. We have also checked all other reported results and found no further transcription errors.


W1: Human Preference Statistics

I noticed that the TaoUGC results for ReVRA in Table 1 exactly match the memory-free configuration in Table 6, rather than the proposed full IEM configuration. This raises questions about which system was actually evaluated. Also, the reported human preferences for brute-force and ReVRA (53%, 50%, and 3% ties) do not sum to 100%. I think these inconsistencies need clarification.

Response: The reported percentages overlap because the same 3% ties were counted toward both methods. Reporting mutually exclusive outcomes resolves this ambiguity: 50% prefer brute-force, 47% prefer ReVRA, and 3% are ties.

In the user study, 3% of cases were judged to have comparable perceptual quality. These cases were included in both preference percentages: the reported 53% for brute-force comprised 50% exclusive preferences plus 3% ties, while the reported 50% for ReVRA comprised 47% exclusive preferences plus the same 3% ties. We have adopted a clearer explain to avoid ambiguity.

Revision: We have revised the statement in Section 4.5 to make the counting explicit: “Brute-force and ReVRA were the exclusive preference in 50% and 47% of cases, respectively, with 3% ties.”

AuthorsOpenReview ↗

Response to Reviewer 3BEJ (Part 2/3)

W2&Q2: Evaluation of Normalized QA-based Methods

Appendix A.6 averages CLIPIQA, MUSIQ, DOVER, and negative NIQE without normalization. Since MUSIQ has a much larger numerical range, this baseline is effectively dominated by MUSIQ. Moreover, Table 5 compares the DPO-trained evaluator against untuned VLMs, making it difficult to distinguish the benefit of pairwise comparison from preference fine-tuning. Interestingly, the untuned VLMs perform substantially worse than the metric-based baseline on KVQ.

Why are the IQA/VQA metrics averaged without normalization in Appendix A.6? Could the authors clarify whether normalized or calibrated metric-based comparisons were evaluated?

Response: Even through the QA-based's accuracy may be improved by e.g., normalization, our DPO-trained evaluator remains more accurate on both datasets, because aggregating differentmetrics still may not cover the variance of the degradations.

The original QA-based method followed the calculation used by existing agents[1] to maintain consistency with their evaluation protocol. We agree that directly averaging metrics with different numerical ranges can give MUSIQ disproportionate influence. We therefore normalize the four metrics, reverse the direction of NIQE, and then average them with equal weights:

Saver=14[CLIPIQA−μCσC+DOVER−μDσD+MUSIQ−μMσM−NIQE−μNσN]S_{\text{aver}}=\frac{1}{4}\left[\frac{CLIPIQA-\mu_C}{\sigma_C}+\frac{DOVER-\mu_D}{\sigma_D}+\frac{MUSIQ-\mu_M}{\sigma_M}-\frac{NIQE-\mu_N}{\sigma_N}\right], where μ\mu and σ\sigma denote the mean and standard deviation of each metric over the entire test set. The comparison is shown below:

DatasetQwen3-VLGemini3.1-proQA-based(Raw)QA-based(Normalized)Ours
KVQ62.06%64.25%92.51%96.82%97.84%
VideoLQ63.91%62.96%64.40%61.71%74.36%

Normalization improves QA-based accuracy on KVQ from 92.51% to 96.82%, reducing its gap to our evaluator to 1.02 percentage points. On VideoLQ, accuracy decreases from 64.40% to 61.71%, while our evaluator achieves 74.36%, yielding a gap of 12.65 percentage points. These no-reference metrics measure different aspects of visual quality using different standards, and their scores can vary widely across video content and degradation types. Normalization puts their scores on similar scales, but does not make them reflect human preferences in the same way. A fixed combination of these metrics may therefore fail to capture how different degradations affect perceived quality across videos. Our pairwise comparison directly judges which video looks better, avoiding the need to aggregate these different metrics, while DPO aligns the judgments with human preferences.

Revision: We have replaced the original QA-based results in Table 5 with the normalized results and clarified the calculation and normalization procedure in Appendix A.6.

AuthorsOpenReview ↗

Response to Reviewer 3BEJ (Part 3/3)

W2&Q3: Ablation Separating Pairwise Comparison from DPO Fine-tuning

Appendix A.6 averages CLIPIQA, MUSIQ, DOVER, and negative NIQE without normalization. Since MUSIQ has a much larger numerical range, this baseline is effectively dominated by MUSIQ. Moreover, Table 5 compares the DPO-trained evaluator against untuned VLMs, making it difficult to distinguish the benefit of pairwise comparison from preference fine-tuning. Interestingly, the untuned VLMs perform substantially worse than the metric-based baseline on KVQ.

Could the authors provide an ablation separating pairwise comparison from DPO fine-tuning? I think this is important for validating the claimed advantage of the comparison-based evaluator.

Response: Preference fine-tuning improves both evaluation strategies, yet our trained pairwise comparison model remains more accurate when both use the same preference data. This supports the combination of pairwise comparison and preference learning beyond the benefit of additional training data alone.

To distinguish these contributions, we add absolute scoring as a control using the same Qwen3-VL backbone and human-annotated video pairs. Specifically, we evaluate both pairwise comparison and absolute scoring before preference fine-tuning, then train their counterparts using DPO and RL2R[2], respectively. The results are shown below:

Assessment StrategyAdditional TrainingKVQVideoLQ
Absolute scoring×\times63.20%55.42%
Relative comparison×\times62.06%63.91%
Absolute scoring√\surd80.96%62.97%
Relative comparison√\surd97.84%74.36%

Before preference fine-tuning, pairwise comparison achieves slightly higher accuracy overall: the two strategies perform similarly on KVQ, while pairwise comparison leads by 8.49 percentage points on VideoLQ. As the reviewer notes, the untuned VLM underperforms QA-based on KVQ, highlighting the importance of preference fine-tuning. Crucially, when both strategies use the same preference training data, our pairwise comparison model still outperforms absolute scoring by 16.88 and 11.39 percentage points on KVQ and VideoLQ, respectively. These results clarify both contributions: preference fine-tuning substantially strengthens the VLM’s quality assessment capability, and both additional preference training and relative comparison contribute to improving visual quality comparison accuracy.

Revision: We have added this ablation and its analysis to Table 9 in Appendix A.6 of the revised manuscript.


[1] Yingjie Zhou. 20252025. Q-agent: Quality-driven chain-of-thought image restoration agent through robust multimodal large language model. Arxiv.

[2] Tianhe Wu, et al. 20252025. Reasoning-Induced Image Quality Assessment via Reinforcement Learning to Rank, Neurips.

AuthorsOpenReview ↗

Further Discussion with Reviewer 3BEJ

Dear Reviewer 3BEJ,

Thank you for your time and for your thoughtful and constructive comments. We have considered your suggestions carefully and respond to them below.

  1. We revise the TaoUGC results in Table 1 and clarify that they correspond to the final system with the IEM enabled.
  2. We clarify how the human preference statistics are computed and revise the relevant statements in Sec. 4.5 to avoid ambiguity.
  3. We provide an evaluation against normalized QA-based metrics, replace the original QA-based results in Table 5, and clarify the computation in Appendix A.6. We further explain why normalization alone does not align different metrics with human preferences.
  4. We extend the ablation separating the benefits of pairwise comparison from those of DPO fine-tuning, and add it to Appendix A.6.

As the discussion phase draws to a close, we hope our responses have addressed your concerns. Please let us know if any further questions remain. We sincerely appreciate your time and feedback.

Best regards,

The Authors.

Reviewer_zn1yOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ReVRA, an agentic framework for restoring videos with compound degradations. At each stage, a planning VLM examines the current video and proposes candidate restoration toolboxes. A DPO-finetuned evaluation VLM compares their outputs, and the preferred result becomes the input to the next stage. A skip option prevents unnecessary processing, while an embedding-based, online-updated memory provides planning priors. The paper also introduces TaoUGC, a benchmark containing 150 real-world UGC clips. Experiments across seven datasets, tool-aligned comparisons, and component ablations demonstrate the framework’s effectiveness in improving perceptual restoration quality.

Critical strengths
  1. Stage-wise toolbox reselection is a useful design for compound-degradation restoration. Restoration tools may simultaneously resolve multiple degradations or introduce new artifacts, making the initial plan unsuitable for subsequent states. ReVRA addresses this by re-examining intermediate videos and grounding decisions in observed execution outcomes. Figure 7 illustrates how dynamic adaptation and skipping can avoid over-processing, while Table 4 provides empirical support for stage-wise replanning.

  2. The experimental evaluation provides several complementary forms of evidence. The seven datasets cover synthetic and real-world degradations. Table 1 shows substantial improvements on DVD and CRVD, with additional reference-based validation in Table 7. The tool-aligned comparison in Table 3 helps reduce confounding from different tool configurations. Furthermore, the human-preference agreement results in Table 5 support the practical value of the trained video comparator. Together, these experiments establish the framework’s usefulness beyond a conceptual system demonstration.

Critical weaknesses
  1. The distinctions from existing agents and the contributions of individual components could be established more clearly. AgenticIR already incorporates reflection and rescheduling, while VQ-Jarvis combines retrieval, trained video comparison, and step-wise operator selection. The contribution should therefore be positioned more specifically around stage-wise toolbox-type reselection and online memory updates. Although the ablations provide initial support, comparisons with VQ-Jarvis are limited, and Table 5 does not fully separate the benefits of relative comparison from those of additional preference training.

  2. Some experimental protocols and reported results require clarification. Appendix A.6 averages quality metrics with substantially different numerical scales without describing normalization, which may affect the QA-based baseline. The TaoUGC results in Table 1 match the memory-free row of Table 6, whereas the full IEM configuration reports different results. Finally, using only the first 30 frames of each video limits the conclusions regarding temporal robustness over longer sequences.

Questions for authors
  1. Could you clarify how stage-wise toolbox reselection differs from the scheduling mechanisms in AgenticIR and VQ-Jarvis? Which existing results most directly demonstrate the benefit of this distinction?

  2. Could you report final restoration performance using the same pairwise-comparison procedure before and after DPO? Would the QA-based results in Table 5 support the same conclusions after normalizing the constituent quality metrics?

  3. Please clarify the memory initialization settings used for the TaoUGC results in Tables 1 and 6 and explain their correspondence. Please also specify the prototype update rule and whether video processing order materially affects the results.

5 comments
AuthorsOpenReview ↗

Response to Reviewer zn1y (Part 1/4)

We thank the reviewer for recognizing the novelty of our stage-wise toolbox reselection, the breadth of our evaluation, and the practical value of this work. Below we clarify how ReVRA differs from existing agents and further validate our evaluation VLM by ii separating the benefits of relative comparison from those of preference training, iiii benchmarking against normalized QA-based methods, and iiiiii verifying that this training improves final restoration quality. We also verify and correct the TaoUGC results in Table 1.


W1&Q1: Distinctions from existing Agents

The distinctions from existing agents and the contributions of individual components could be established more clearly. AgenticIR already incorporates reflection and rescheduling, while VQ-Jarvis combines retrieval, trained video comparison, and step-wise operator selection. The contribution should therefore be positioned more specifically around stage-wise toolbox-type reselection and online memory updates. Although the ablations provide initial support, comparisons with VQ-Jarvis are limited, and Table 5 does not fully separate the benefits of relative comparison from those of additional preference training.

Could you clarify how stage-wise toolbox reselection differs from the scheduling mechanisms in AgenticIR and VQ-Jarvis? Which existing results most directly demonstrate the benefit of this distinction?

Response: ReVRA differs from existing agents through stage-wise toolbox-type reselection and online memory updates.

For planning, ReVRA’s stage-wise toolbox-type reselection dynamically adjusts the restoration strategy based on actual intermediate results. AgenticIR only reschedules the order of the remaining toolboxes after restoration failure, without changing their types, while VQ-Jarvis fixes both toolbox types and order. Table 4 directly supports this design: DOVER improves from 0.752 under static planning to 0.795 with ReVRA.

Regarding memory, AgenticIR and VQ-Jarvis use static memory, with VQ-Jarvis directly reusing retrieved restoration pipelines for easier cases, whereas ReVRA retrieves a planning prior from the current video state and updates IEM online using execution outcomes. Table 6 directly validates the benefit of online memory updates: with the planning VLM retained in both settings, online IEM Row6Row 6 outperforms static memory Row4Row 4 on all four restoration quality metrics.

Table 2 further compares ReVRA with VQ-Jarvis on YouHQ40, where VQ-Jarvis has publicly reported results. ReVRA outperforms it on all three commonly reported metrics: CLIPIQA, MUSIQ, and DOVER. Results on the other evaluated datasets are unavailable because VQ-Jarvis has not released its source code, preventing us from faithfully reproducing those results.

Revision: 

  1. We will revise ReVRA's positioning in Section 1 from “ReVRA decomposes the restoration into multiple stages, each performing a complete plan–execute–evaluate workflow for tool selection and execution” to “ReVRA introduces stage-wise toolbox-type reselection based on observed intermediate results and online memory updates driven by execution outcomes.”

  2. We will also clarify the corresponding distinctions from AgenticIR and VQ-Jarvis in Section 2.2, positioning the contribution more specifically around stage-wise toolbox-type reselection and online memory updates.

AuthorsOpenReview ↗

Response to Reviewer zn1y (Part 2/4)

W1: Separating the Benefits of Relative Comparison and Preference Training

The distinctions from existing agents and the contributions of individual components could be established more clearly. AgenticIR already incorporates reflection and rescheduling, while VQ-Jarvis combines retrieval, trained video comparison, and step-wise operator selection. The contribution should therefore be positioned more specifically around stage-wise toolbox-type reselection and online memory updates. Although the ablations provide initial support, comparisons with VQ-Jarvis are limited, and Table 5 does not fully separate the benefits of relative comparison from those of additional preference training.

Response: Both additional preference training and relative comparison improve visual quality assessment accuracy.

Table 5 already demonstrates the benefit of additional preference training by comparing relative comparison before and after training. To distinguish this benefit from that of relative comparison, we additionally evaluate Qwen3-VL using absolute scoring. Specifically, we train an absolute scoring model using RL2R [3] on the same Qwen3-VL backbone and human-annotated video pairs. The results are shown below.

Assessment StrategyAdditional TrainingKVQVideoLQ
Absolute scoring×\times63.20%55.42%
Relative comparison×\times62.06%63.91%
Absolute scoring√\surd80.96%62.97%
Relative comparison√\surd97.84%74.36%

Without additional preference training, relative comparison achieves slightly higher accuracy overall: the two strategies perform similarly on KVQ, while relative comparison leads by 8.49 percentage points on VideoLQ. With additional preference training, both strategies improve substantially, and the advantage of relative comparison becomes more pronounced: it outperforms absolute scoring by an average of 14 percentage points across the two datasets. These results demonstrate that both additional preference training and relative comparison contribute to more accurate visual quality assessment.

Revision: We have added this ablation and its analysis to Table 9 in Appendix A.6 of the revised manuscript.


W2&Q2: Evaluation of Normalized QA-based Methods

Some experimental protocols and reported results require clarification. Appendix A.6 averages quality metrics with substantially different numerical scales without describing normalization, which may affect the QA-based baseline. The TaoUGC results in Table 1 match the memory-free row of Table 6, whereas the full IEM configuration reports different results. Finally, using only the first 30 frames of each video limits the conclusions regarding temporal robustness over longer sequences.

Could you report final restoration performance using the same pairwise-comparison procedure before and after DPO? Would the QA-based results in Table 5 support the same conclusions after normalizing the constituent quality metrics?

Response: Our evaluation VLM remains more accurate than normalized QA-based methods on both datasets.

For QA-based evaluation, we followed prior work[4] by averaging the constituent metrics without normalization. Following the reviewer’s suggestion, we normalize each metric before averaging and reverse the direction of NIQE, the normalized average score SaverS_{\text{aver}} is computed as follows:

Saver=14[CLIPIQA−μCσC+DOVER−μDσD+MUSIQ−μMσM−NIQE−μNσN]S_{\text{aver}}=\frac{1}{4}\left[\frac{CLIPIQA-\mu_C}{\sigma_C}+\frac{DOVER-\mu_D}{\sigma_D}+\frac{MUSIQ-\mu_M}{\sigma_M}-\frac{NIQE-\mu_N}{\sigma_N}\right],

where μ\mu and σ\sigma are the mean and standard deviation of each metric over the entire test set. After normalization, the accuracy of the QA-based method is compared with that of other methods, as shown in the table below:

DatasetQwen3-VLGemini3.1-proQA-based(Raw)QA-based(Normalized)Ours
KVQ62.06%64.25%92.51%96.82%97.84%
VideoLQ63.91%62.96%64.40%61.71%74.36%

Normalization improves QA-based accuracy on KVQ by 4.31 percent points, but reduces it by 2.69 percent points on the more challenging VideoLQ. Our evaluation VLM remains ahead by 1.02 and 12.65 percent points, respectively. Thus, normalization does not change the conclusion: correcting the numerical scales alone does not consistently align QA-based judgments with human preferences.

Revision:  We have replaced the original QA-based results in Table 5 with the normalized results and clarified the calculation in Appendix A.6.

AuthorsOpenReview ↗

Response to Reviewer zn1y (Part 3/4)

W2&Q2: Effect of DPO Training for Evaluation VLM on Final Restoration Performance

Some experimental protocols and reported results require clarification. Appendix A.6 averages quality metrics with substantially different numerical scales without describing normalization, which may affect the QA-based baseline. The TaoUGC results in Table 1 match the memory-free row of Table 6, whereas the full IEM configuration reports different results. Finally, using only the first 30 frames of each video limits the conclusions regarding temporal robustness over longer sequences.

Could you report final restoration performance using the same pairwise-comparison procedure before and after DPO? Would the QA-based results in Table 5 support the same conclusions after normalizing the constituent quality metrics?

Response: The before-and-after DPO comparison shows that better alignment with human preferences improves final restoration performance.

To examine whether the benefit of DPO also carries through to final restoration performance, we keep all other components and settings fixed on TaoUGC and replace only the DPO-trained evaluation VLM with its untuned Qwen3-VL counterpart, retaining the same pairwise-comparison procedure.

NIQE↓CLIPIQA↑MUSIQ↑DOVER↑
Before DPO(Qwen3-VL)5.4450.47661.870.818
After DPO(evaluation VLM)5.1910.49364.060.821

DPO improves all four restoration metrics, including CLIPIQA from 0.476 to 0.493 and MUSIQ from 61.87 to 64.06. The improvement therefore extends beyond evaluation accuracy: more accurate preference judgments lead to better final restoration performance.

Revision: We have added the before-and-after DPO ablation in Appendix A.6.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Response to Reviewer 86yM (Part 1/2)

We thank the reviewer for the review and for recognizing the novelty of our multi-stage self-corrective reflection framework and the Instance-aware Evolving Memory mechanism. In response to the concerns and suggestions raised by the reviewer, we have added all requested ablations and analyses, including the ablation on the number of reflection stages, the stability analysis of the evaluation VLM, and the analysis of error propagation and correction along the pipeline. We organize our responses below accordingly.


W1&Q1: Ablation on the Number of Reflection Stages

Missing ablations isolating the necessity of multi‑stage reflection. Although the paper argues convincingly that fixed pipelines fail, it does not provide a direct comparison between single‑stage ReVRA and multi‑stage ReVRA. Such an ablation would strengthen the claim that reflection—not just better memory or better evaluation—is the key driver of performance.

Can you provide ablations comparing 1‑stage, 2‑stage, and full multi‑stage ReVRA to quantify the benefit of reflection?

Response: The full multi-stage ReVRA outperforms both its 1-stage and 2-stage variants, confirming that reflection is the key driver of performance.

Following the reviewer's suggestion, we compare two ablation variants with fewer reflection stages—1-stage and 2-stage—against the full multi-stage ReVRA used in our main experiments on VideoLQ. In the 1-stage setting, reflection is performed only at the final stage, where it determines the entire toolbox chain; in the 2-stage setting, reflection occurs at the first and last stages, determining the first toolbox and the remaining chain, respectively. Overall performance improves as the number of reflection stages increases: relative to 1-stage, DOVER rises by 0.011 for 2-stage and by 0.055 for full multi-stage. This confirms that reflection is the key driver of performance.

Reflection stagesNIQE↓CLIPIQA↑MUSIQ↑DOVER↑
1-stage4.0960.42956.890.740
2-stage3.9270.42457.050.751
multi-stage (ReVRA)3.8900.44959.300.795

Revision: We supplement the ablation on the number of reflection stages in Appendix A.9.


W2&Q2: Stability of Evaluation VLM

Reliability of pairwise VLM evaluation is under‑analyzed. The evaluation VLM is fine‑tuned via DPO, but the paper does not analyze its stability across different degradation types or its sensitivity to subtle artifacts. Since evaluation drives stage transitions and memory updates, more evidence of robustness would be valuable.

How stable is the DPO‑tuned evaluation VLM across different degradation types and content domains?

Response: The DPO‑tuned evaluation VLM is stable across different degradation types and content domains.

We further break down KVQ dataset by scene content and distortion type, grouping the videos and reporting the selection accuracy of each group, against a normalized QA-based baseline.

Different content domains: Based on scene content, we partition the KVQ test set into 8 overlapping categories. The table below reports each method's accuracy per category. Our DPO-tuned evaluation VLM achieves higher overall accuracy and stays above 90% in every category, whereas the QA-based baseline varies widely: it reaches 99.10% on film clips but drops to 81.63% on text-dominant videos. This indicates that individual IQA metrics favor certain content types, and that no unified anchor is available to eliminate such biases under simple averaging.

Live-actionVFXCG animationStill imagestext-dominantStage performancefilm clipsScreen recordingTotal
Evaluation VLM98.27%92.31%100%93.26%100%100%100%95.12%97.84%
QA-based98.02%98.22%94.29%95.51%81.63%94.74%99.10%98.78%96.82%

Different degradation types: KVQ also covers the four degradation types common in UGC; the table below reports accuracy per type. Our DPO-tuned evaluation VLM stays above 97% on all four, whereas the QA-based baseline remains less stable, with a standard deviation of 2.71 compared with 0.98 for ours.

Compression artifactsBlurLow resolutionNoiseTotal
Evaluation VLM97.60%99.03%100%100%97.84%
QA-based97.32%96.59%92.45%100%96.82%

These results show that our DPO-tuned evaluation VLM is stable across both degradation types and content domains. We attribute this to two factors. First, the evaluation VLM is pretrained on large-scale video data and therefore brings a general visual prior to quality judgment. Second, our DPO training data covers diverse content and distortion types. As a result, the evaluation VLM shows no pronounced sensitivity to content or degradation type.

AuthorsOpenReview ↗

Response to Reviewer 86yM (Part 2/2)

Q3: Analysis of Error Propagation and Correction

How sensitive is ReVRA to errors in early stages—does a wrong toolbox selection propagate, or does reflection reliably correct it?

Response: ReVRA prevents 91% of incorrect tool selections in the early stages, avoiding the propagation of wrong toolbox selections.

To evaluate the sensitivity of ReVRA to errors in early stages, we conduct additional perturbation experiments. Specifically, on the GoPro dataset containing only blur and compression degradations, we inject perturbations to force ReVRA to invoke an incorrect denoising toolbox at Stage 1, observing whether it can prevent error propagation or correct the error via reflection. The results are reported as follows, where the proportion of videos that actually execute the denoising operation is denoted as the Error Decision Rate. Compared with the unperturbed baseline, the perturbed ReVRA avoids 91% of errors and their propagation at Stage 1 by invoking the skip operation. It is worth emphasizing that the error decision rate of the unperturbed ReVRA remains 0. Consequently, the perceptual metrics degrade only slightly while remaining superior to other state-of-the-art methods listed in Table 1. This indicates that ReVRA successfully prevents most errors from arising and propagating via reflection, thereby validating the robustness of the proposed system.

PerturbationError Decision RateNIQECLIPIQAMUSIQDOVER
√\surd9%4.8720.26546.170.666
×\times0%4.6970.28647.260.679

Revision: We supplement the analysis of this error propagation and correction experiment in Appendix A.8.

AuthorsOpenReview ↗

Further Discussion with Reviewer 86yM

Dear Reviewer 86yM,

We appreciate your insightful and constructive review. Your suggestions have helped us strengthen the experimental evidence and clarify the role of reflection in ReVRA. Below are our point-by-point responses:

  1. We extend the ablation on the number of reflection stages and add it to Appendix A.9. It shows that full multi-stage ReVRA outperforms the 1-stage and 2-stage, supporting the benefit of repeated reflection.
  2. We evaluate the stability of the evaluation VLM across different degradation types and content domains. It achieves above 97% accuracy on all tested degradation types and above 90% on all tested content domains, supporting its reliability across these groups.
  3. We experimentally analyze error propagation and correction in ReVRA, added to Appendix A.8. The experiment shows that ReVRA rejects 91% of forced denoising outputs through skip, with only a 0.013 decrease in final DOVER, demonstrating limited sensitivity to this early-stage perturbation.

As the discussion phase draws to a close, we hope these results address your concerns and would appreciate your thoughts. Please let us know if any further questions remain. Thank you for your time and feedback.

Best regards,

The Authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.