Research map / Vision-language and multimodal / Multimodal LLMs
Q-Insight: Understanding Image Quality via Visual Reinforcement Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Q-Insight, a GRPO-trained Qwen2.5-VL-7B model that jointly learns image quality score regression and degradation perception from limited labels with verifiable rewards, generalizing zero-shot to comparison reasoning.
Problem
Image quality assessment with MLLMs has either produced uninterpretable numerical scores or relied on large SFT datasets of textual descriptions; a unified interpretable model that both scores and reasons about quality is lacking.
Approach
Applies GRPO with binary verifiable rewards (score within threshold epsilon, degradation class and level correctness) plus a format reward, jointly training score regression and degradation perception so the tasks mutually benefit, and transfers zero-shot to pairwise comparison reasoning.
Key results
- Score regression: average PLCC 0.806 / SRCC 0.783 across seven datasets, vs DeQA-Score 0.786 / 0.765; gains on OOD sets such as CSIQ (0.870 PLCC vs 0.787).
- Degradation perception: average distortion accuracy 0.9277 vs AgenticIR 0.5998, with level accuracy 0.4796 vs 0.1922.
- SRIQA-Bench overall accuracy 82.80% for trained Q-Insight vs 82.40% for A-FINE and 62.96% for DepictQA; zero-shot version reaches 75.51%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add degradation perception training to score regression (ablation) | Q-Insight · Score regression across KonIQ, SPAQ, KADID, PIPAL, LiveW, AGIQA, and CSIQ | PLCC and SRCC: KonIQ: 0.918 -> 0.933 PLCC; 0.895 -> 0.916 SRCC; AVG.: 0.765 -> 0.806 PLCC; 0.739 -> 0.783 SRCC | Helped |
| Add score regression training to degradation perception (ablation) | Q-Insight · Degradation perception test | Deg. Acc. and Lev. Acc.: Average Deg. Acc.: 0.8960 -> 0.9277; Average Lev. Acc.: 0.3492 -> 0.4796 | Helped |
| Vary the score reward threshold epsilon from 0.15 to 0.50 (ablation) | Q-Insight · Score regression across KonIQ, SPAQ, KADID, PIPAL, LiveW, AGIQA, and CSIQ | PLCC and SRCC: KonIQ: 0.918/0.900 to 0.933/0.916; SPAQ: 0.903/0.903 to 0.907/0.905; KADID: 0.749/0.744 to 0.742/0.736; PIPAL: 0.443/0.4 | No clear effect |
| Use Q-Insight instead of DeQA | Qwen-2.5-VL-7B-Instruct · KonIQ | PLCC: 0.953 vs 0.933 | Hurt |
| Use Q-Insight instead of DeQA | Qwen-2.5-VL-7B-Instruct · SPAQ, KADID, PIPAL, LiveW, AGIQA, and CSIQ | PLCC and SRCC: SPAQ PLCC: 0.895 -> 0.907; KADID PLCC: 0.694 -> 0.742; PIPAL PLCC: 0.472 -> 0.486; CSIQ PLCC: 0.787 -> 0.870 | Helped |
| Use Q-Insight instead of AgenticIR | Qwen-2.5-VL-7B-Instruct · Degradation perception test | Average Deg. Acc. and Lev. Acc.: Deg. Acc.: 0.5998 -> 0.9277; Lev. Acc.: 0.1922 -> 0.4796 | Helped |
| Use Q-Insight instead of AgenticIR | Qwen-2.5-VL-7B-Instruct · Null degradation category | Deg. Acc.: 0.9339 -> 0.7603 | Hurt |
| Train Q-Insight on labeled comparison pairs rather than use it zero-shot | Qwen-2.5-VL-7B-Instruct · SRBench | Overall-Acc: 75.51% -> 82.80% | Helped |
| Use trained Q-Insight instead of A-FINE | Qwen-2.5-VL-7B-Instruct · SRBench | Overall-Acc: 82.40% -> 82.80% | Helped |
Surprising findings
- Training with score labels alone gives poor perception of fine degradations such as JPEG compression, which joint degradation training fixes.
- Q-Insight is weaker than DeQA-Score on in-domain KonIQ, while winning on OOD datasets.
Limitations
- Focuses primarily on natural images; AI-generated images and videos not covered.
- Score reward is binary with a threshold, so it discards finer MOS information.
- Slightly weaker than DeQA-Score on in-domain KonIQ.
Experimental setup
Datasets: KonIQ DQ-495K DiffIQA SPAQ KADID PIPAL LIVE-Wild AGIQA CSIQ SRIQA-Bench
Benchmarks: KonIQ SPAQ KADID PIPAL LIVE-Wild AGIQA CSIQ SRIQA-Bench
Metrics: PLCC SRCC accuracy (degradation class and level) pairwise accuracy
Baselines: DeQA-Score Q-Align C2Score Qwen-SFT MUSIQ CLIP-IQA+ ManIQA NIMA NIQE BRISQUE AgenticIR DepictQA
Builds on: GRPO DeepSeek-R1 Qwen2.5-VL Q-Align DeQA-Score
Closest papers by meaning
- VisualQuality-R1: Reasoning-Induced Image Quality Assessment via Reinforcement Learning to Rank (NeurIPS 2025 · similarity 0.92)
- Reasoning as Representation: Rethinking Visual Reinforcement Learning in Image Quality Assessment (ICLR 2026 · similarity 0.92)
- Q-Probe: Scaling Image Quality Assessment to High Resolution via Context-Aware Agentic Probing (NeurIPS 2026 · similarity 0.91)
- Refine-IQA: Multi-Stage Reinforcement Finetuning for Perceptual Image Quality Assessment (AAAI 2026 · similarity 0.89)
- Self-Evolving Vision-Language Models for Image Quality Assessment via Voting and Ranking (ICLR 2026 · similarity 0.88)
- UARE: A Unified Vision-Language Model for Image Quality Assessment, Restoration, and Enhancement (CVPR 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.