PaperlayerView exact revision on arXiv

Source-linked AI summary

Q-Insight: Understanding Image Quality via Visual Reinforcement Learning

Weiqi Li, Xuanyu Zhang, Shijie Zhao, Yabin Zhang, Junlin Li, Li Zhang, Jian Zhang

arXiv:2503.22679v2cs.CV

TL;DR

IQA needs models that capture perceptual quality, degradation, and comparisons while balancing precise scores with interpretable reasoning. Q-Insight applies GRPO to jointly train score regression and degradation perception from limited labels. It outperforms existing methods across quality-assessment tasks and generalizes zero-shot to comparison reasoning, while remaining focused primarily on natural images.

  • Problem

    Existing IQA methods and MLLM-based approaches do not jointly provide comprehensive visual understanding, precise scoring, and interpretable reasoning without extensive annotation.

  • Method

    Q-Insight uses GRPO to jointly optimize score regression and degradation perception with specialized rewards and limited MOS and degradation labels.

  • Results

    Q-Insight consistently outperforms existing IQA metrics and SFT-driven language models across datasets and tasks, with zero-shot generalization to image comparison reasoning.

  • Takeaways & Limitations

    The framework combines numerical quality scoring with interpretative reasoning for comprehensive image quality understanding across multiple assessment and comparison tasks.

  • Takeaways & Limitations

    Q-Insight focuses primarily on natural images; extending it to AI-generated images and videos remains future work.

Abstract

from arXiv · show

Image quality assessment (IQA) focuses on the perceptual visual quality of images, playing a crucial role in downstream tasks such as image reconstruction, compression, and generation. The rapid advancement of multi-modal large language models (MLLMs) has significantly broadened the scope of IQA, moving toward comprehensive image quality understanding that incorporates content analysis, degradation perception, and comparison reasoning beyond mere numerical scoring. Previous MLLM-based methods typically either generate numerical scores lacking interpretability or heavily rely on supervised fine-tuning (SFT) using large-scale annotated datasets to provide descriptive assessments, limiting their flexibility and applicability. In this paper, we propose Q-Insight, a reinforcement learning-based model built upon group relative policy optimization (GRPO), which demonstrates strong visual reasoning capability for image quality understanding while requiring only a limited amount of rating scores and degradation labels. By jointly optimizing score regression and degradation perception tasks with carefully designed reward functions, our approach effectively exploits their mutual benefits for enhanced performance. Extensive experiments demonstrate that Q-Insight substantially outperforms existing state-of-the-art methods in both score regression and degradation perception tasks, while exhibiting impressive zero-shot generalization to comparison reasoning tasks. Code will be available at https://github.com/lwq20020127/Q-Insight.

1 Introduction

Existing IQA methods struggle to capture comprehensive visual quality and generalize out of distribution, while MLLM-based approaches trade precise scoring against interpretable reasoning. Q-Insight uses GRPO with joint score regression and degradation perception to achieve broad quality-understanding capabilities from limited labels.

  • Motivation: Traditional and deep-learning IQA methods often miss global visual quality and struggle with out-of-distribution generalization.Traditional methods rely on handcrafted comparisons or statistics, while learned models face OOD challenges.
  • Motivation: MLLM-based IQA methods typically provide either adaptable numerical scores without sufficient interpretability or detailed descriptions without precise scoring.This leaves unified numerical scoring and descriptive reasoning unresolved.
  • Approach: Q-Insight applies Group Relative Policy Optimization to jointly optimize score regression and degradation perception using limited MOS and degradation labels.The framework uses verifiable score, degradation classification, and intensity perception rewards.
  • Approach: Q-Insight is designed as a reasoning-style MLLM for comprehensive image quality understanding rather than score prediction alone.Its target capabilities include image description, aesthetic and compositional evaluation, degradation perception, and comparative reasoning.
  • Results: Extensive experiments report consistent gains over existing IQA metrics and SFT-driven language models, including zero-shot generalization to image comparison reasoning.Figure 1 highlights improved PLCC performance, especially on out-of-domain datasets such as CSIQ.

2 Related Work

Related work spans score-based IQA, description-based MLLM assessment, and reinforcement-learning approaches for reasoning and vision-language alignment. Q-Insight combines these directions by applying GRPO to visual quality understanding with multi-task training.

  • Score-based IQA: Score-based IQA includes full-reference methods that compare distorted images with references and non-reference methods that estimate quality without references.Methods range from traditional metrics such as SSIM to deep-learning-derived quality priors.
  • Description-based IQA: Description-based IQA uses MLLM knowledge to produce qualitative assessments and improve interpretability across perceptual and comparative tasks.Examples include Q-Bench, Q-Instruct, Co-Instruct, and DepictQA-family methods.
  • Reinforcement learning: Reinforcement learning improves language-model reasoning through feedback-driven refinement, including rule-based GRPO without supervised fine-tuning.Prior vision-language applications include visual grounding, medical reasoning, and multimodal reasoning.
  • Position of Q-Insight: Q-Insight is presented as the first work integrating reinforcement learning into a foundational visual quality-understanding model.It jointly trains multiple tasks and reports mutually beneficial effects among them.

3 Methodology

Q-Insight applies GRPO to jointly train score regression and degradation perception, using task-specific rewards to encourage structured reasoning and robust image-quality understanding from limited annotations.

  • 3.1 Preliminaries: GRPO samples response groups, computes relative advantages from their task-specific rewards, and updates the policy while clipping policy ratios and applying KL regularization.The approach compares responses within each group rather than relying on an explicit critic model.
  • 3.3 Multi-Task Group Relative Policy Optimization: Q-Insight jointly optimizes score regression and degradation perception from image-question inputs, generating responses with explicit reasoning steps.The two tasks use score and degradation-related rewards during multi-task policy optimization.
  • 3.3 Multi-Task Group Relative Policy Optimization: A shared format reward requires reasoning and answer tags plus a JSON-like answer structure, enabling consistent parsing across tasks.Responses receive format reward 1 only when all stated formatting conditions are satisfied.
  • 3.3 Multi-Task Group Relative Policy Optimization: The score-regression reward assigns 1 when the predicted score falls within threshold ϵ of the ground-truth MOS and 0 otherwise.Setting ϵ to 0 reduces the reward to exact-answer matching.
  • 3.3 Multi-Task Group Relative Policy Optimization: Degradation perception uses separate classification and intensity rewards, with intensity receiving credit only when both distortion class and level exactly match the ground truth.The task uses degradation labels spanning five distortion categories and five severity levels.
  • 3.4 Data Construction: The overall reward combines format and task-specific terms, allowing Q-Insight to switch between tasks and jointly optimize them using limited annotated labels.The framework is constructed for score regression and degradation perception and is intended to support robust reasoning and perception.

4 Experiments

Experiments evaluate Q-Insight on score regression, distortion perception, multi-task training, threshold robustness, and image comparison reasoning. Across these settings, the model shows strong out-of-distribution performance, reciprocal task benefits, and zero-shot comparison capability.

  • Experimental Setup: Q-Insight uses diverse IQA datasets spanning in-the-wild, synthetic, model-processed, and AI-generated images for score regression evaluation.KonIQ provides approximately 7000 training images, while MOS serves as the ground-truth score.
  • Score Regression: Q-Insight’s score-regression example combines a numerical rating with reasoning about text legibility, lighting, focus, and image composition.One example predicts 3.70 against a ground-truth score of 3.63.
  • Score Regression: Approximately 0.02 PLCC and SRCC improvements on OOD datasets show Q-Insight consistently outperforming baselines across nearly all benchmarks, despite slightly underperforming DeQA-Score on KonIQ.The comparison includes handcrafted, deep-learning, and MLLM-based IQA methods trained under the stated evaluation setup.
  • Distortion Perception: Q-Insight outperforms AgenticIR across nearly all distortion categories while identifying distortion types with a single query rather than sequential queries.Performance is slightly lower for the null category, possibly because the reasoning mechanism is more sensitive to potential distortions.
  • Ablation Studies: Joint training improves both tasks: degradation perception strengthens score regression, while score regression improves degradation perception across distortion types.The gains are especially notable on synthetic and model-generated distortion datasets for score regression.
  • Image Comparison Reasoning: Q-Insight generalizes zero-shot to reference-based and non-reference-based image comparison reasoning, outperforming DepictQA in content understanding and degradation perception.Training on 5k labeled comparison pairs further boosts comparison performance, and the model surpasses score-based and description-based methods on reported metrics.

5 Conclusion

The paper concludes that Q-Insight is a GRPO-based model for comprehensive image quality understanding using limited labeled data. It combines numerical scoring with interpretative reasoning, achieves strong results across tasks, and remains primarily evaluated on natural images.

  • 5 Conclusion: Q-Insight jointly optimizes score regression and degradation perception with limited labeled data, combining numerical accuracy with interpretative reasoning.The framework is presented as an alternative to extensive textual annotations and purely numerical scoring.
  • 5 Conclusion: Experiments report consistent outperformance of existing state-of-the-art methods across datasets and tasks, alongside zero-shot generalization and strong comparison reasoning.The conclusion characterizes Q-Insight as a unified model for scoring, perception, comparison, and reasoning.
  • 5 Conclusion: Q-Insight focuses primarily on natural images, leaving extension to AI-generated images and videos for future exploration.This is the stated limitation of the current study’s scope.
  • A Designed Prompts: The prompt design uses a shared system prompt and task-specific prompts for score regression, degradation perception, and comparative reasoning.Inputs include images paired with questions, with MOS or distortion class and severity as task labels.
  • A Designed Prompts: The score-regression prompt requests a quality rating from 1 to 5, while the degradation prompt classifies distortion type and severity in structured JSON-like output.The comparison prompts select between Image A, Image B, or Similar depending on the scenario.

B.1 Ablation study of the threshold ϵ

The threshold ablation evaluates score-regression performance across several ϵ values and finds robust, stable results without careful threshold tuning.

  • B.1 Ablation study of the threshold ϵ: The score reward is granted when prediction error is less than ϵ, allowing scores to vary within an acceptable margin.The threshold controls the tolerance used in score-regression reward computation.
  • B.1 Ablation study of the threshold ϵ: Robust and stable performance across ϵ = 0.15, 0.25, 0.30, 0.40, and 0.50 indicates that Q-Insight does not depend on meticulous threshold tuning.Table B reports PLCC/SRCC values across KonIQ, SPAQ, KADID, PIPAL, LiveW, AGIQA, and CSIQ.

B.2 Effect of multi-task training

Joint multi-task training improves Q-Insight’s ability to perceive detailed image degradations and supports more accurate image quality assessment. The qualitative examples illustrate this benefit through degradation-sensitive scoring explanations.

  • Joint training enables Q-Insight to perceive detailed degradations such as pixel-level artifacts, improving overall image quality assessment accuracy.The ablation describes this as a qualitative benefit of combining score regression with degradation perception.
  • Without joint training, the model’s image-quality reasoning can identify broad visual properties but may miss degradation cues that affect scoring.The supplied scoring examples mention lighting, colors, noise, blur, pixelation, sharpness, and clarity as factors in ratings.

C.1 Datasets

The datasets combine paired image-quality comparisons with real-world super-resolution evaluation. SRIQA-Bench specifically tests generalization when only degraded low-resolution references are available.

  • DiffIQA contains approximately 180k reference–test image pairs generated by diffusion-based enhancement methods, with human preference labels from triplet-based comparisons.Each reference image is paired with multiple test images, producing roughly 180k comparison pairs.
  • SRIQA-Bench includes 100 low-resolution reference images, each enhanced by 10 distinct super-resolution methods and evaluated through exhaustive pairwise human comparisons.Each comparison is annotated by at least 10 annotators.
  • Because SRIQA-Bench provides no ground-truth high-resolution references, models must judge perceptual quality from degraded low-resolution images alone.The benchmark measures robustness under imperfect reference conditions using ranking accuracy and correlation metrics.

C.2 Reward Design and Training Details

The comparison-reasoning training procedure assigns rewards to model responses and combines them into an overall reward for policy optimization. Training uses AdamW with a decaying learning rate and fixed compute settings.

  • The image-comparison reward evaluates whether a response correctly identifies the superior image or declares the pair comparable.The comparison task expects one of these three outcomes.
  • The overall reward for each response combines task-specific reward components with a format reward.The supplied training description explicitly states that the overall reward is calculated from these components.
  • Training uses AdamW with an initial learning rate of 1 × 10^-6 linearly decaying to 1 × 10^-9 over 8 epochs.The total batch size is 32, and training uses 16 NVIDIA A100 GPUs for approximately 20 hours.

C.3 More Qualitative Results

Additional qualitative comparisons show Q-Insight recognizing subtle distortions and making image-quality judgments that align with the stated reference-based comparison outcomes. The examples include both choosing a superior image and identifying comparable outputs.

  • C.3 More Qualitative Results: After training on only 5k comparison pairs, Q-Insight surpasses DepictQA qualitatively by recognizing subtle image details and distortions more effectively.The passage presents this as producing more accurate reference-based comparison outcomes.
  • C.3 More Qualitative Results: For brick-building images, Image B is judged higher quality because it retains better detail and texture definition despite slight darkening, whereas Image A is blurred.The associated MOS values are Image A 1.20 and Image B 1.00, while the qualitative judgment favors Image B.
  • C.3 More Qualitative Results: Another comparison marks both outputs as similar because both are blurry, lack detail, and show no discernible fidelity difference relative to the reference.The stated ground truth for this example is Similar.
  • C.3 More Qualitative Results: For the red-train example, Q-Insight selects Image B because its cleaner appearance and lower noise provide a better balance of clarity and detail than Image A.Image A is described as more vibrant and detailed but slightly grainy; the stated ground truth is Image B.
Loading 2503.22679v2…