VisualQuality-R1: Reasoning-Induced Image Quality Assessment via Reinforcement Learning to Rank

We are deeply grateful for the reviewer’s thoughtful insights and thorough evaluation of our manuscript. Please find what below our detailed point-to-point responses to all the comments of this reviewer. We hope our responses can address this reviewer's concerns.

[W1 & Q2] Regarding Computational Overhead

Thanks for the constructive comment. We would like to clarify that during inference, we generate only one response per image (i.e., K=1), and the corresponding latency and memory usage metrics have been reported in our responses to Reviewer Cn1F [W4] and Reviewer NF6n [W1 & Q1]. Please refer to those sections for details.

During training, we set the number of generations in GRPO to K=6. To better illustrate the efficiency trade-offs versus performance gains, we provide a detailed comparison of training cost and model performance under different generation numbers K=4,5,6. As shown in the table below, increasing K leads to a moderate and manageable increase in both GPU memory usage and training time. Meanwhile, the PLCC improves consistently from 0.811 at K=4 to 0.814 at K=6, suggesting that additional generations bring incremental yet stable gains.

\begin{array}{lccc} \hline & \text{K=4} & \text{K=5} & \text{K=6} \newline \hline \text{Memory} & 58\text{G per GPU} & 62\text{G per GPU} & 68\text{G per GPU} \newline \text{Time} & 3\text{h}39\text{min} & 4\text{h} 23\text{min} & 4\text{h} 45\text{min} \newline \text{Averaged PLCC} & 0.811 &0.813 & 0.814 \newline \hline \end{array}

This empirical evidence further motivates our proposed direction of developing more efficient reasoning strategies, such as sample-adaptive generation or rationale distillation, as discussed in the section of Limitations and Future Directions in the main paper. We will include this clarification in the revised manuscript.

[W2 & Q1] Regarding Prompt Rigidity

Thanks for the valuable and constructive feedback. The limitation of using a fixed prompt has been honestly acknowledged in the section of Limitations and Future Directions (lines 252–256 in the main paper). To take a step forward, we designed a small set of scenario-specific prompts based on the original template in Table 1. For the detailed prompt settings, kindly refer to our responses to Reviewer NF6n [W2] . For the corresponding experimental results and analysis, please refer to our responses to Reviewer Cn1F [W5 & Q1]. In realistic scenarios, using adaptive prompts at inference time leads to noticeable improvements, highlighting the potential of our method to better adapt to distortion-specific contexts.

Regarding the observation that Fig. C shows scores for stylized images lack nuance (e.g., rating digital art "4.00" without style sensitivity), we thank the reviewer for pointing this out. We will include this as a representative failure case in the revised manuscript to illustrate the model’s limited sensitivity to artistic style when using fixed prompt. This also underscores the importance of designing scenario-specific prompts to better capture nuanced aspects such as artistic style, which we identify as a promising direction for future work.

[W3] Regarding Baseline Depth

Thanks for raising this important point. We agree that including comparisons with recent VLM-based methods is valuable. In fact, our main paper already includes several recent VLM-based NR-IQA baselines that are open-sourced and reproducible, such as LIQE, Q-Align, and DeQA-Score. These represent state-of-the-art approaches that leverage supervised fine-tuning on vision-language backbones.

As for LLaVA-IQA, to the best of our knowledge, it is only briefly mentioned in a recent non-IQA-specific work ([1] VistaDream, ICCV 2025). To address this concern, we implemented a LLaVA-based baseline under the same evaluation protocol. As shown in the table below, LLaVA-IQA performs notably worse in terms of SRCC than VisualQuality-R1, particularly on distortion-focused datasets such as BID, CLIVE, and Deblurring. This underscores the challenge of adapting general-purpose VLMs to NR-IQA without task-specific training or reinforcement-based optimization. We will include this comparison in the revised manuscript.

\begin{array}{l|cccc|cccc|c} \hline \text{Method} & \text{BID} & \text{CLIVE} & \text{KonIQ} & \text{SPAQ} & \text{Deblurring} & \text{Super-Resolution} & \text{Dehazing} & \text{Image Generation} & \text{Avg} \newline \hline \text{LLaVA-IQA} & 0.446 & 0.401 & 0.437 & 0.672 & 0.620 & 0.633 & 0.488 & 0.612 &0.539 \newline \text{VisualQuality-R1} & 0.791 & 0.747 & 0.829 & 0.875 & 0.838 & 0.751 & 0.599 & 0.776 & 0.776 \newline \hline \end{array}

[1] VistaDream: Sampling multiview consistent images for single-view scene reconstruction, ICCV 2025

[W4] Regarding Societal Impact

Thanks for raising this important point. While we initially assessed the societal impact as minimal (§9), we acknowledge that VisualQuality-R1 could be misused in automated content moderation to suppress images deemed “low-quality.” Although our model is designed to support image enhancement and retrieval, not censorship, we agree that such risks warrant explicit discussion.

We will address this in the revised manuscript by highlighting the potential misuse and recommending safeguards, such as human-in-the-loop review and transparency about model limitations. We appreciate the reviewer for bringing attention to this important issue.

[Q3] Regarding Implausible Score

Thanks for pointing out this issue. The score of 0.20 in Fig. B indeed falls outside the instructed [1, 5] range, and we sincerely acknowledge this as an instruction-following failure, which is a common limitation observed in large-scale language models.

In our experiments, we noticed that during the later training epochs (e.g., epochs 9 to 10), the model occasionally produced scores slightly below 1 or above 5. When training continued beyond epoch 10, these anomalies persisted, but they did not noticeably affect overall SRCC and PLCC performance, as both of which are relative ranking metrics.

We believe there are two plausible reasons for this behavior:

• Since the model is optimized with a learning-to-rank objective, it first learns to correctly rank images near the boundary scores (e.g., around 1 and 5). After this initial convergence, when it encounters samples that appear significantly worse than previous low-quality images (or significantly better than previous high-quality ones), it may exaggerate the relative differences by assigning scores below 1 or above 5 to maximize the fidelity-based reward.
• More subtly, during policy optimization, the trade-off between reward maximization and KL regularization may not perfectly constrain the output head to obey soft output bounds. In rare cases, the model may exploit this slack to push scores slightly beyond the expected range.

We will explicitly document this behavior in the revised manuscript and treat it as a known limitation. One potential remedy is to impose a soft or hard clipping mechanism to enforce outputs within the range without disrupting gradient flow.

[Q4] Regarding The Extension to FR-IQA

Thanks for pointing this out. We fully agree that extending VisualQuality-R1 to FR-IQA is a promising research direction, as also outlined in our Conclusion and Discussion section. In this rebuttal, we take a preliminary step toward this goal by adapting our model to FR-IQA settings and conducting initial experiments.

We begin by modifying the prompt to include explicit reference comparison instructions. Specifically, we adopt the following prompt structure:

''... Here is the question: The first image serves as the reference for the second image. Please compare the second image against the reference and evaluate its overall visual quality. Provide a quality score as a float between 1.00 and 5.00, ...''

We fine-tuned VisualQuality-R1 on KADID-10K in the FR setting and evaluated it on three representative FR benchmarks: Deblurring, Super-Resolution, and Dehazing. The SRCC results are as follows. (Note: ''-FR'' uses both reference and distorted images; ''-NR'' uses only the distorted image.)

\begin{array}{l|ccc} \hline \text{Method} &\text{Deblurring} & \text{Super-Resolution} & \text{Dehazing} \newline \hline \text{Qwen2.5-VL-7B-NR} & 0.820 & 0.603 & 0.458 \newline \text{VisualQuality-R1-NR} & 0.872 & 0.825 & 0.651 \newline \hline \text{Qwen2.5-VL-7B-FR} & 0.751 & 0.488 & 0.389 \newline \text{VisualQuality-R1-FR} & 0.869 & 0.767 & 0.581 \newline \hline \end{array}

As shown in the above table, VisualQuality-R1-FR, fine-tuned with RL2R, significantly outperforms the zero-shot Qwen2.5-VL-7B in the full-reference setting. This demonstrates the effectiveness of our method in adapting VLM to the FR setting.

However, we also note that VisualQuality-R1-FR still underperforms its NR counterpart. Similarly, in the zero-shot setting, Qwen2.5-VL-7B shows a notable drop in performance under FR. Upon further qualitative inspection, we found that in the FR setting, the model’s reasoning outputs frequently fail to capture subtle differences between the distorted and reference images. In some cases, the model even incorrectly concludes that the two images are identical, despite the visible degradations. These results suggest that reasoning over paired inputs in the FR setting is substantially more challenging and requires targeted learning strategies, which we leave for future work.

These experiments serve as an early feasibility study toward extending VisualQuality-R1 to FR-IQA. We will include the results and analysis in the revised manuscript.

Dear Reviewer 5jDt,

Many thanks for your time in reviewing our paper and your constructive comments. We have submitted the point-to-point responses. We appreciate if you could let us know whether your concerns have been addressed, and we are happy to answer any further questions.

Best regards,

Authors of paper #1789

Thank you for the author's response. This is a great piece of work, and I recommend Borderline accept

We sincerely appreciate this reviewer's insightful feedback and careful review of our manuscript. Please find what below our detailed point-to-point responses to all the comments of this reviewer. We hope our responses can address this reviewer's concerns.

[W1] Regarding Qualitative Reasoning Comparison

Thanks for the constructive suggestion. We agree that qualitative comparisons can provide valuable insight into the interpretability of model outputs. In the revised version of our manuscript, we will include additional visual examples comparing the reasoning outputs of VisualQuality-R1 with those of Q-Insight and other representative methods to better showcase the strengths of our approach.

Additionally, we have conducted a user study to assess the alignment of VisualQuality-R1’s rationales with human perception of image quality. Please refer to our responses to Reviewer NF6n [W3 & Q2] for more details. The results show that 87.5% image-text pairs were judged as aligned. This provides evidence that the reasoning generated by VisualQuality-R1 is largely consistent with human perceptual judgments, despite the absence of explicit supervision at the rationale level.

[W2] Regarding Failure Case Demonstration

We sincerely thank this reviewer for this insightful suggestion. As also indicated by Reviewer 5jDt [Q1 & Q3], our model indeed exhibits some typical failure cases. For example, in the supplementary material, the second image in Figure B produces a score that falls outside the instructed range, and the sixth image in Figure C lacks style-specific reasoning in its explanation. We will include more failure cases in the revised manuscript. These examples will help illustrate scenarios where the model struggles and highlight directions for future improvement.

[W3] Regarding Continuous Reward

Thanks for the insightful comment. To further investigate the practical significance of continuous rewards over binary rewards, we conducted an additional ablation study under a multi-dataset training setting, which differs from the single-dataset setup used in the main paper. Specifically, we combined four heterogeneous IQA datasets: KADID, TID2013, CLIVE, and SPAQ. The SRCC results are shown in the table below.

\begin{array}{l|cccc|cccc|c} \hline \text{Method} & \text{BID} & \text{CLIVE} & \text{KonIQ} & \text{SPAQ} & \text{Deblurring} & \text{Super-Resolution} & \text{Dehazing} & \text{Image Generation} & \text{Avg} \newline \hline \text{Binary Reward} & 0.822 & 0.817 & 0.844 & 0.911 & 0.849 & 0.743 & 0.614 & 0.752 & 0.794 \newline \text{Continuous Reward} & 0.829 & 0.824 & 0.845 & 0.916 & 0.856 & 0.740 & 0.662 & 0.757 & 0.804 \newline \hline \end{array}

The results consistently show that continuous reward yields higher PLCC scores on 7 out of 8 tasks, leading to a noticeable +1.0% absolute gain on average. Notably, the improvements are more pronounced on challenging cross-domain settings like Dehazing (+4.8%) and Deblurring (+0.7%), where binary rewards tend to saturate. This suggests that continuous fidelity rewards provide more informative and fine-grained supervision signals, helping the policy model better differentiate subtle perceptual quality differences during reinforcement learning.

These findings align with our hypothesis that continuous rewards better reflect the nuanced nature of human perception, which is inherently more graded than binary. As such, even if individual score gaps may seem small, the consistent directional gain across diverse tasks and datasets affirms the practical benefit and robustness of the continuous reward design.

[W4] Regarding Multi-Dataset Training

Thanks for the constructive comment. To address the concern regarding limited dataset diversity in our multi-dataset training setup, we conducted an additional ablation study by combining four heterogeneous IQA datasets (KADID, TID2013, CLIVE, and SPAQ) into a joint training set. This setup introduces greater variation in distortion types, image content, and rating scales, providing a more rigorous evaluation of model generalization. The SRCC results are presented in the table below.

\begin{array}{l|cccc|cccc|c} \hline \text{Training Dataset} & \text{BID} & \text{CLIVE} & \text{KonIQ} & \text{SPAQ} & \text{Deblurring} & \text{Super-Resolution} & \text{Dehazing} & \text{Image Generation} & \text{Avg} \newline \hline \text{KADID} & 0.790 & 0.750 & 0.830 & 0.875 & 0.838 & 0.756 & 0.598 & 0.775 & 0.777 \newline \text{KADID-SPAQ} & 0.811 & 0.811 & 0.855 & 0.913 & 0.845 & 0.752 & 0.588 & 0.754 & 0.791 \newline \text{KADID-TID2013-CLIVE-SPAQ} & 0.829 & 0.824 & 0.845 & 0.916 & 0.856 & 0.740 & 0.662 & 0.757 & 0.804 \newline \hline \end{array}

Compared to training solely on KADID (Avg = 0.777) or KADID-SPAQ (Avg = 0.791), training on all the four datasets further improves the average SRCC to 0.804, demonstrating that VisualQuality-R1 benefits from the increased diversity and volume of training data. Importantly, the performance gains are not isolated to a single dataset or distortion type. For instance, we observe improvements on BID and Dehazing datasets, which are known to pose generalization challenges. This confirms that our RL2R framework remains effective under broader and more heterogeneous training conditions. We will include this experiment in the revised manuscript.

Dear Reviewer JU4b,

Many thanks for your time in reviewing our paper and your constructive comments. We have submitted the point-to-point responses. We appreciate if you could let us know whether your concerns have been addressed, and we are happy to answer any further questions.

Best regards,

Authors of paper #1789

We sincerely thank this reviewer for the insightful feedback and the thorough evaluation of our manuscript. Please find what below our detailed point-to-point responses to all the comments of this reviewer. We hope our responses can address this reviewer's concerns.

[W1 & Q1] Regarding Computational Cost Analysis

Thanks for the constructive comment. The concern regarding the computational overhead of the proposed model is fully acknowledged. To address this issue, we have added a quantitative comparison of inference time, GPU memory usage, and FLOPs across several representative NR-IQA models. These results will be included in the revised manuscript.

\begin{array}{l|c|c|c|c} \hline \text{Method} & \text{Parameters} & \text{Inference Time (BS=1)} & \text{Per-GPU Inference Memory} & \text{TFLOPs} \newline \hline \text{UNIQUE} & 22.32\text{M} & 0.02\text{s} & 1.62\text{G} & 0.029 \newline \text{MUSIQ} & 27.13\text{M} & 0.05\text{s} & 1.69\text{G} & 0.026 \newline \text{MANIQA} & 135.75\text{M} & 0.03\text{s} & 2.13\text{G} & 0.217 \newline \text{LIQE} & 151.28\text{M} & 0.03\text{s} & 2.15\text{G} & 0.131 \newline \text{Q-Align} & 8.20\text{B} & 0.14\text{s} & 17.1\text{G} & 1.98 \newline \text{DeQA-Score} & 8.20\text{B} & 0.11\text{s} & 17.1\text{G} & 1.98 \newline \text{Q-Insight} & 8.29\text{B} & 2.72\text{s} & 17.6\text{G} & 8.71 \newline \text{VisualQuality-R1} & 8.29\text{B} & 2.34\text{s} & 17.6\text{G} & 7.74 \newline \hline \end{array}

As shown in the table above, VisualQuality-R1 incurs higher inference cost than discriminative models such as MUSIQ and MANIQA, as well as small-scale VLMs like LIQE, primarily due to the large model backbone (Qwen2.5-VL-7B). However, this computational overhead reflects a deliberate design choice aimed at enabling fine-grained, human-aligned reasoning, uncertainty-aware ranking under the Thurstone model, and consistent generalization across diverse distortion scenarios (see Table 2 in the main paper). With the growing popularity and adoption of large-scale VLMs, the increased computational overhead is acceptable.

[W2] Regarding Prompt Robustness

We appreciate the reviewer’s constructive comment regarding the robustness of fixed prompts. In response, we investigate prompt robustness from two complementary perspectives: 1) introducing new scenario-specific prompt content, and 2) performing controlled perturbations of the original prompt wording. The results will be included in the revised manuscript.

1) Scenario-Specific Prompt: As a preliminary step toward addressing this issue, we conducted a small-scale investigation by introducing scenario-specific prompt variants at inference time, without modifying the training configuration. Instead of using a single fixed instruction, we customized the prompt to match the distortion scenario of each test image (e.g., deblurring, super-resolution, dehazing). Specifically, we modified the segment in Table 1 from: ''Rate the overall image visual quality. The rating should be a float between 1 and 5…'' to: ''Rate the overall image visual quality. [scenario-specific prompt] The rating should be…''.

The [scenario-specific prompt] is dynamically inserted based on the distortion type. Below are examples of the specialized prompts:

• Realistic scenario: The image may contain naturally occurring distortions from digital imaging pipelines, such as noise, blur, compression artifacts, or color inaccuracy. Evaluate the overall perceived quality as a human would perceive it in everyday photography settings.
• Deblurring: This image may have undergone a deblurring process. Evaluate the overall visual quality, including sharpness, clarity, and naturalness. Consider whether the deblurring contributes to a visually pleasing result, or if it introduces artifacts, unnatural textures, or over-sharpening effects. Focus on the image as a whole and assess how perceptually high-quality it appears.
• Super-Resolution: This image was enhanced by a super-resolution model. Rate its visual quality based on detail, sharpness, coherence, and realism. Watch for artifacts like over-sharpening, unnatural textures, or structural errors.
• Dehazing: This is a dehazed image. Assess the overall visual quality after haze removal, focusing on whether the result appears natural, clear, and visually pleasing. Your evaluation should reflect the image’s clarity, realism, and aesthetic coherence as a whole.
• Image Generation: This image is generated by an AI model. Evaluate its overall perceptual quality based on both structural coherence and aesthetic appeal. For aesthetics, consider factors such as composition, color harmony, lighting, visual balance, and artistic style.

Rather than altering the full prompt, we introduced a minimal scenario-specific prefix while keeping the core instruction unchanged. Please refer to our responses to Reviewer Cn1F [W5 & Q1] for detailed results and analysis. Without modifying training, such prompts bring clear gains for imaging-related distortions, but offer limited improvement for complex processing-related ones, possibly because the model has already captured key cues.

2) Controlled Perturbations: We apply prompt perturbations at inference time by randomly introducing minor modifications to the instruction in Table 1, including word substitutions, deletions, and the insertion of special characters. The experimental SRCC results are summarized in the table below. Across all eight datasets, we observe negligible performance differences, suggesting that the model is robust to such minor variations in prompt formulation.

\begin{array}{l|cccc|cccc|c} \hline \text{Method} & \text{BID} & \text{CLIVE} & \text{KonIQ} & \text{SPAQ} & \text{Deblurring} & \text{Super-Resolution} & \text{Dehazing} & \text{Image Generation} & \text{Avg} \newline \hline \text{Original} & 0.790 & 0.750 & 0.830 & 0.875 & 0.838 & 0.756 & 0.598 & 0.775 & 0.777 \newline \text{Perturbed} & 0.791 & 0.747 & 0.829 & 0.875 & 0.838 & 0.751 & 0.599 & 0.776 & 0.776 \newline \hline \end{array}

[W3 & Q2] Regarding Empirical Substantiation of Rationales

We sincerely appreciate the reviewer’s constructive suggestions. We acknowledge that, to date, there is no well-established metric for accurately evaluating text–image alignment at the low-level quality dimension such as blur, noise, or artifacts, unlike high-level semantic similarity. Hence, we conducted a human study to assess the perceptual alignment of the model’s rationales. We curated a set of 40 test images that were not seen during training, including both natural photographs and outputs from various enhancement algorithms such as deblurring, super-resolution, and AI-generated content. For each image, we collected the model-generated rationale and asked human evaluators to assess whether the reasoning accurately reflects perceptually important quality factors, such as sharpness, contrast, and noise.

Each image-text pair was independently evaluated by 15 domain experts in low-level vision. Following a hard-labeling protocol, we considered a pair to be aligned if at least 10 out of 15 raters deemed the rationale accurate. Each participant took approximately 30 minutes to complete the evaluation.

The results show that 35 out of 40 image-text pairs (87.5%) were judged as aligned. This provides evidence that the reasoning generated by VisualQuality-R1 is largely consistent with human perceptual judgments, despite the absence of explicit supervision at the rationale level. These results will be included in the revised version of the paper.

[W4 & Q3] Regarding Overclaim Generalization

Thanks for the thoughtful comments. This point touches the core design philosophy of RL2R.

Rescaling MOS Scores: We clarify that our method does not require rescaling MOS scores to a 1–5 range during training. Rescaling is applied only in Q-Insight’s multi-dataset setup (see footnote, page 7 of our main paper). RL2R’s output is constrained to the 1–5 range solely to align with Q-Insight but this does not imply that input score rescaling is required during training.

As shown in Equation 2, our model relies only on relative comparisons between image quality scores. The continuous reward is based on pairwise preferences, independent of score scale. This relative information can be directly derived from MOS scores in any range (e.g., 0–100 or 1–5), enabling VisualQuality-R1 to train on datasets without perceptual scale realignment.

This design also enables VisualQuality-R1 to utilize datasets that contain only weak or relative supervision, such as BAPPS or DiffIQA, where fine-grained MOS scores are unavailable. Compared to regression-based methods like Q-Insight, which depend heavily on MOS values and require careful normalization, our RL2R formulation is fundamentally more scalable and robust to diverse labeling conditions.

Generalization to Unlabeled Settings: Previous work [42] has shown that unlabeled data can be leveraged via gMAD, where image pairs are selected such that a weaker model rates them similarly while a stronger model detects significant perceptual differences. Building on this, we suggest using VisualQuality-R1 as the student and identifying teacher models with superior performance in certain distortion types or content. gMAD can help mine image pairs from unlabeled data where the teacher sees differences that VisualQuality-R1 misses. These pairs can then be used to fine-tune VisualQuality-R1 and enhance its generalization without human labels.

[42] Active fine-tuning from gMAD examples improves blind image quality assessment, TPAMI 2021

Dear Reviewer NF6n,

Many thanks for your time in reviewing our paper and your constructive comments. We have submitted the point-to-point responses. We appreciate if you could let us know whether your concerns have been addressed, and we are happy to answer any further questions.

Best regards,

Authors of paper #1789

Reviewer NF6n,

Thanks a lot for your initial review. Please read authors' rebuttal and provide your reply/discussion ASAP before the deadline. Thanks

Your area chair

We are truly grateful for this reviewer's valuable insights and careful evaluation on our manuscript. Please find what below our detailed point-to-point responses to all the comments of this reviewer. We hope our responses can address this reviewer's concerns.

[W1] Regarding the Base Design

We appreciate the reviewer’s thoughtful concern. However, we respectfully cannot agree that the base design alone is insufficient. To clarify the motivation and demonstrate the effectiveness of our design, which is formulated as pure relative scoring through RL2R, we elaborate on two key aspects below.

1. As shown in Table 4, VisualQuality-R1, which uses only relative scoring without additional components, achieves SRCC of 0.770 and PLCC of 0.799, comparable to Q-Insight’s 0.766 and 0.803. This demonstrates that the core design is already competitive with existing strong methods, validating its effectiveness and soundness.

2. Continuous rewards, adaptive variance estimation, and probability averaging are enhancements based on the our relative comparison framework. As shown in Table 4, all variants are built upon this foundation, demonstrating its modularity and extensibility. This flexibility is a result of our framework’s mathematical grounding and statistical expressiveness.

[W2] Regarding the Generalizability and Robustness in Single-Dataset Scenarios

VisualQuality-R1 and Q-Insight share the same vision-language backbone, Qwen2.5-VL-7B, and they are fully fine-tuned from identical pre-trained weights. The key difference lies in the learning approach: Q-Insight treats IQA as an absolute regression task with binary rewards, while our model adopts a relative formulation using RL2R, applying continuous fidelity-based rewards and uncertainty-aware comparisons based on the Thurstone model. A 1% gain without architectural changes or pretraining highlights RL2R’s strength, and consistent improvements across individual datasets confirm its robustness and generalizability in single-dataset settings. We argue that the improvement made by our our method is not marginal.

More importantly, RL2R is originally designed for multi-dataset training, where scale misalignment is a major challenge. As shown in Table 4, simply applying Q-Insight’s regression-based objective to multi-dataset training leads to a significant performance drop. In contrast, our RL2R approach not only maintains performance but achieves substantial gains across datasets, further validating its effectiveness in more challenging settings.

[W3] Regarding Batch Construction

Thanks for the insightful comment. We fully agree that batch construction is important for relative comparison models. We analyzed the impact of per-GPU batch size and sample quality, with SRCC results shown in the table below.

\begin{array}{l|cccc|cccc|c|c} \hline \text{Method} & \text{BID} & \text{CLIVE} & \text{KonIQ} & \text{SPAQ} & \text{Deblurring} & \text{Super-Resolution} & \text{Dehazing} & \text{Image Generation} & \text{Avg} & \text{Time} \newline \hline \text{Q-Insight} & 0.784 & 0.761 & 0.806 & 0.872 & 0.831 & 0.724 & 0.601 & 0.749 & 0.766 & 5\text{h} \newline \text{BS=4} & 0.793 & 0.764 & 0.826 & 0.880 & 0.840 & 0.764 & 0.602 & 0.769 & 0.780 & 8\text{h} \newline \text{BS=6} & 0.793 & 0.768 & 0.830 & 0.881 & 0.841 & 0.755 & 0.607 & 0.764 & 0.780 & 6\text{h} \newline \text{BS=8 (Ours)} & 0.790 & 0.750 & 0.830 & 0.875 & 0.838 & 0.756 & 0.598 & 0.775 & 0.777 & 5\text{h} \newline\hline \text{Same Content} &0.744 &0.708 &0.796 &0.857 &0.830 &0.765 &0.536 &0.778 &0.752 & 5\text{h} \newline \text{Similar Quality} &\text{fail} &\text{fail} &\text{fail} &\text{fail} &\text{fail} &\text{fail} &\text{fail} &\text{fail} &\text{fail} &- \newline \hline \end{array}

Impact of Batch Size: We experimented with per-GPU batch sizes of 4, 6, and 8. We observed that reducing the batch size results in only marginal performance gains (with SRCC improvements within 0.003), but it leads to a substantial increase in training time. For example, the training time increases from 5 hours with BS=8 to 8 hours with BS=4. This inefficiency becomes even more critical under multi-dataset training, where the overall data scale is significantly larger. Therefore, we choose BS=8 to ensure both effectiveness and scalability.

Impact of Sampling Strategies (Sample Quality): Our default approach samples images randomly from a single dataset to form each batch. To further investigate the effect of sample quality, we examine two alternative strategies:

1. Same Content Sampling: We sample batches in which all images share the same content but differ in distortion levels, a strategy applicable only to FR-IQA datasets like KADID-10K. As shown in the table (row: Same Content), this leads to lower performance (SRCC: 0.752), indicating that restricting content diversity weakens generalization. We attribute this to the context-sensitive nature of IQA, where diverse content and distortions offer richer visual cues. In contrast, uniform content may cause overfitting to limited artifacts, which runs counter to our model’s emphasis on semantic reasoning across heterogeneous image pairs.

2. Similar Quality Sampling: We test a more extreme strategy by sorting images by predicted quality and sampling consecutive ones to reduce intra-batch variance (row: Similar Quality). This leads to training collapse, with the model assigning nearly identical scores to all inputs. The failure highlights the importance of intra-batch diversity for generating effective learning signals in ranking-based frameworks. Diverse content and distortion levels are essential for maintaining visual contrast and enabling robust gradient updates in RL2R. Additionally, exposing vision-language models to a range of quality cases supports stable convergence under continuous reward supervision.

[W4] Regarding Computational Cost

We sincerely thank this reviewer for the suggestion. As mentioned in Lines 170–171 of the main paper, our training was conducted on 16 NVIDIA A100 GPUs and took approximately 5 hours. To better address the reviewer’s concern, we provide a detailed summary of memory consumption and runtime for both training and inference in the table below.

\begin{array}{lccc} \hline \text{Computational Cost} & \text{Training (BS=8, K=6, 10 epochs)} & \text{Inference (BS=32)} & \text{Inference (BS=1)} \newline \hline \text{Memory} & 68\text{G per GPU} & 23.8\text{G per GPU} & 17.6\text{G per GPU} \newline \text{Time} & 5\text{h} & 5.62\text{s (One-Shot)} & 2.34\text{s (One-Shot)} \newline \hline \end{array}

The table summarizes the memory and time costs for both training and inference, all evaluated on images with a resolution of 512×384. Training with batch size 8, K equals 6, and 10 epochs takes about 5 hours and uses 68 GB of GPU memory since the multi-response sampling and pairwise ranking are performed with a 7B model. In inference, the one-shot setting refers to generating a single response per prompt. Under this setting, inference with batch size 32 takes 5.62 seconds and uses 23.8 GB of memory, while batch size 1 takes 2.34 seconds and uses 17.6 GB, which fits within the memory capacity of commonly available GPUs such as the RTX 3090 (24 GB). Please also refer to our responses to Reviewer NF6n [W1 & Q1] for more details. We will include this information in the revised manuscript.

[W5 & Q1] Regarding Single Structured Prompt

We sincerely thank this reviewer for the insightful suggestion. We have explicitly listed this as a limitation of our method and highlighted application-aware prompt adaptation as a future direction (please refer to lines 252–256 in our main paper). To further investigate the potential benefit of prompt specialization, we conducted an additional ablation where we kept the training process unchanged and varied the prompt only during inference. Please refer to our responses to Reviewer NF6n [W2] for the full set of scenario-specific prompts. The updated SRCC results are shown in the table below.

\begin{array}{l|cccc|cccc|c} \hline \text{Method} & \text{BID} & \text{CLIVE} & \text{KonIQ} & \text{SPAQ} & \text{Deblurring} & \text{Super-Resolution} & \text{Dehazing} & \text{Image Generation} & \text{Avg} \newline \hline \text{Fixed Prompt} & 0.790 & 0.750 & 0.830 & 0.875 & 0.838 & 0.756 & 0.598 & 0.775 & 0.777 \newline \text{Diverse Prompt} & 0.795 & 0.767 & 0.840 & 0.887 & 0.837 & 0.742 & 0.588 & 0.778 & 0.779 \newline \hline \end{array}

The above results reveal several interesting patterns, analyzed as follows:

Imaging-Related Distortions (BID, CLIVE, KonIQ, SPAQ): All the four datasets exhibit consistent improvements. This suggests that adding simple textual cues helps the model contextualize low-level distortions such as blur or noise more effectively.

Processing-Related Distortions (Deblurring, Super-Resolution, Dehazing, Image Generation): These types of distortions often involve complex trade-offs, such as the trade-offs between fidelity and perceptual quality. In such cases, a short prompt may not provide sufficient guidance, and the model might already capture most of the relevant cues from the image itself, making the additional textual input either unnecessary or potentially confusing.

In this ablation, we only vary the inference-time prompt without retraining the model. Greater gains may be possible by incorporating prompt specialization into training (e.g., via prompt tuning or reward-aware generation), which we leave for future work.

Dear Reviewer Cn1F,

Many thanks for your time in reviewing our paper and your constructive comments. We have submitted the point-to-point responses. We appreciate if you could let us know whether your concerns have been addressed, and we are happy to answer any further questions.

Best regards,

Authors of paper #1789