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

• NR-IQA is practical when pristine references are unavailable, but generalization to unseen distortions remains unresolved; Q-Insight [21] relies on dataset-specific rewards and distortion classification.These constraints reflect Q-Insight's framing of visual quality as an absolute regression target.Source: §1 · p. 2
• VLMs can contextualize NR-IQA by combining low-level distortions with high-level attributes such as aesthetics and semantics, enabling nuanced quality descriptions with stronger generalization.Source: §1 · p. 2
• VLM-based SFT requires costly quality-description annotation, risks fitting training-data biases and forgetting pretrained knowledge, and often produces rigid templates; GPT-4o annotation also incurs computational and financial burdens [16].Source: §1 · p. 2
• VisualQuality-R1 assesses diverse distortions, outperforms discriminative NR-IQA models and the reasoning-induced regression method [21], generates human-aligned descriptions, and supports multi-dataset training without perceptual scale realignment.Its descriptions can provide targeted feedback to image-processing algorithms and support fine-grained quality control.Source: §1 · pp. 2–3
• RL2R uses GRPO to generate multiple scores per image pair, compares images with Thurstone probabilities based on score means and sample variances, and rewards estimates using continuous fidelity.This design treats visual quality as inherently relative rather than as an absolute regression target.Source: §1 · p. 2

Regression-based Models

• Regression-based NR-IQA evolved from handcrafted features to end-to-end deep score prediction [3] [39] [54] [52], but often generalizes poorly and requires perceptual-scale realignment across datasets [31].Methods used natural-scene, degradation-specific, or human-vision-inspired features, then standard regression losses or statistical distances.Source: §2 · p. 3

Ranking-based Models

• Ranking-based NR-IQA treats quality as relative, progressing from anchor-based and pairwise methods to end-to-end Thurstone ranking, though variance scaling ambiguity persisted.Fidelity loss outperformed cross-entropy; fixing variance to one enabled active fine-tuning on challenging examples and continual adaptation, alongside margin, correlation, and permutation-probability ranking losses.Source: §2 · p. 3

VLM-based Models

• VLM-based NR-IQA uses multimodal representations for contextual quality assessment, while VisualQuality-R1 shifts from absolute regression to relative ranking.Prior approaches include CLIP multitask adaptation and SFT methods generating scores, distributions, or descriptions; Q-Insight uses RL for reasoning-induced regression but faces dataset-specific reward calibration, auxiliary-classification complexity, and weak generalization to novel distortions. VisualQuality-R1 targets stronger cross-distortion generalization and quality justifications.Source: §2 · p. 3

3 Reasoning-Induced NR-IQA

• VisualQuality-R1 combines reasoning-inducing reinforcement learning with relative quality ranking by integrating the Thurstone model into GRPO.For image pairs, the method generates multiple quality-score responses, each with a reasoning process, and estimates the probability that one image is perceived as better.Source: §3 · pp. 3–4

3.1 VisualQuality-R1 via RL2R

• The fine-tuned VLM produces a 1–5 quality score through step-by-step reasoning, with its prompt specifying tagged thinking and answer outputs.The structured prompt requests a float rounded to two decimal places, with 1 denoting very poor quality and 5 excellent quality.Source: §3.1 · p. 4, Table 1 · p. 5
• GRPO generates K scores per image, and Thurstone comparisons standardize each score difference using sample variances to represent image-specific predictive uncertainty.Using the sample mean for the comparison helps stabilize probability and reward estimates by penalizing outlier predictions.Source: §3.1 · p. 4
• Each score receives a continuous fidelity reward based on human preference, giving RL2R finer ranking supervision than discretized binary rewards.Rewards are standardized within each image’s group to compute relative advantages.Source: §3.1 · pp. 4–5, §3 · p. 4
• The GRPO policy update uses reward-weighted likelihood with clipping and KL regularization to limit destabilizing updates and excessive deviation from the reference policy.The reference policy comes from VLM pre-training, while the previous-epoch policy supplies the sampled reasoning trajectories.Source: §3.1 · p. 5
• VisualQuality-R1 supports multi-dataset training and adaptation without perceptual scale realignment, while aiming to improve generalization and produce contextual textual justifications alongside scores.The paper also presents RL2R as mitigating scalability and overfitting issues associated with SFT-based models.Source: §3.1 · p. 5

4.2 Main Results

• VisualQuality-R1 achieves the best average results, and reasoning-induced models outperform SFT-based counterparts; base Qwen2.5-VL-7B [1] reaches 0.708 SRCC without IQA-specific training.The paper attributes VisualQuality-R1's average performance to RL2R's stronger alignment with human quality perception than regression-based approaches.Source: Table 2 · p. 6, §4.2 · pp. 6–7

4.1 Experimental Setups

• Single-dataset evaluation trains on KADID-10K and tests zero-shot across eight datasets; VisualQuality-R1 fine-tunes Qwen2.5-VL-7B [1] with GRPO [36].Training uses K=6 candidate responses and β=0.04; the setup also compares handcrafted, discriminative deep-learning, and VLM-based IQA models. It runs for 10 epochs on 16 NVIDIA A100 GPUs.Source: §4.1 · pp. 5–6, §4.2 · p. 6

4.3 Ablation Studies

• Reducing GRPO responses from K=6 to four or five has only a marginal effect, trading a small accuracy change for lower computational cost.Other settings are held fixed in this comparison.Source: §4.3 · p. 7, Table 3 · p. 7
• Continuous fidelity rewards generalize better than GRPO's [36] binary reward, while both variants outperform regression-based Q-Insight [21].The comparison is within the same RL2R framework.Source: §4.3 · p. 7, Table 4 · p. 7

4.4 Further Analysis

• Prediction-score variability falls steadily during training across CLIVE, KonIQ-10k, SRIQA-Bench, and AGIQA-3K, indicating increasingly stable assessments.The analysis averages score standard deviations for 20 sampled images from each dataset at successive checkpoints.Source: Figure 3 · p. 8, §4.4 · p. 8

4.3 Ablation Studies

• Averaging quality scores outperforms averaging comparison probabilities, while fixing variances to one degrades performance on nearly all datasets.The findings indicate that sample variances capture pairwise perceptual difficulty and improve comparison reliability and fidelity reward stability.Source: §4.3 · pp. 7–8, Table 4 · p. 7

4.4 Further Analysis

• Across training iterations, VisualQuality-R1 progresses from superficial observations to nuanced reasoning that identifies subtle super-resolution artifacts and calibrates its quality score.On a SwinIR [22] image, its final explanation attributes softness to filtering or object structure, calls the appearance “surreal,” and assigns 3.00; Q-Insight [21] scores it 2.00 and Qwen2.5-VL-7B [1] scores it 4.80.Source: §4.4 · p. 8

Conclusion and Discussion

• RL2R integrates the Thurstone model into GRPO to capture predictive uncertainty, using continuous fidelity rewards for more precise policy-gradient signals.The approach is grounded in the intrinsic relativity of visual quality.Source: §5 · p. 8
• Across reasoning stages, VisualQuality-R1 progresses from overlooking artifacts to detecting stylization, blur, and filtering effects, producing more accurate scores and human-aligned justifications.The example uses an image super-resolved by SwinIR [22].Source: Figure 4 · p. 9
• VisualQuality-R1 surpasses strong discriminative deep learning-based methods and a reasoning-induced baseline, while bridging synthetic and realistic distortion performance gaps.Its contextually rich, human-aligned descriptions support transparency, interpretability, user trust, and downstream tasks such as content filtering, local enhancement prioritization, and quality-aware image retrieval.Source: §5 · p. 9
• VisualQuality-R1 is slow, expensive, memory-hungry, and can amplify early errors; proposed remedies include adaptive reasoning, rationale compression or distillation, and self-consistency sampling.Its fixed prompt also ignores distortion scenario and application context; prompt tuning or dynamic template selection could improve flexibility and accuracy.Source: §5 · p. 9
• Future work could extend VisualQuality-R1 to reference-based assessment using a potentially corrupted reference image [51], and adapt RL2R to other perceptual tasks.Suggested tasks include image aesthetics assessment, human age estimation, and perceptual similarity ranking [9].Source: §5 · p. 9

A Generalization Probing via gMAD Competition

• In gMAD, models attack a defender by finding image pairs it scores similarly but they rate differently; human observers assess these adversarial pairs.The framework targets model strengths and weaknesses without requiring exhaustive ground-truth labeling.Source: A Generalization Probing via gMAD… · p. 14
• On KonIQ-10k [15], VisualQuality-R1 exposes perceptual inconsistencies in Q-Insight’s predictions and resists adversarial challenges in close alignment with human perception.The paper presents this combination of aggressiveness and robustness as evidence of stronger generalization and reliability for real-world IQA.Source: A Generalization Probing via gMAD… · p. 14

B Model Complexity Comparison

• VisualQuality-R1 has higher inference costs than MUSIQ [19], MANIQA [54], and LIQE [59], reflecting its design for fine-grained quality reasoning and robust ratings across diverse distortion scenarios.The comparison is presented in Table 5.Source: B Model Complexity Comparison · p. 14

C Additional Visual Reasoning Examples

• VisualQuality-R1 provides visual reasoning examples under both synthetic and realistic distortion scenarios.The examples are shown in Figures 6 and 7, respectively.Source: C Additional Visual Reasoning Examples · p. 14, Figure 6 · p. 16, Figure 7 · p. 17
• VisualQuality-R1 occasionally produces scores outside [1, 5], but these deviations have negligible impact on overall quality prediction performance.One example assigns a score of 0.2.Source: C Additional Visual Reasoning Examples · pp. 14–15, Figure 6 · p. 16