Research map / Image generation / Autoregressive and latent generation
VisualQuality-R1: Reasoning-Induced Image Quality Assessment via Reinforcement Learning to Rank
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.
Trains Qwen2.5-VL-7B with GRPO and a Thurstone-model ranking objective plus continuous fidelity rewards to produce reasoning-induced no-reference image quality scores with descriptions.
Problem
No-reference image quality assessment (NR-IQA) generalizes poorly across distortion types; SFT-based VLM methods are costly, overfit, and give templated outputs, and regression-based RL methods treat quality as absolute and need dataset-specific reward calibration.
Approach
Reinforcement learning to rank (RL2R): GRPO samples multiple quality scores per image, comparative probabilities between image pairs are computed under the Thurstone model using sample means and variances, and rewards are continuous fidelity measures against MOS-derived preferences.
Key results
- Single-dataset training on KADID-10K: VisualQuality-R1 reaches average SRCC 0.777 vs 0.766 for Q-Insight and 0.708 for base Qwen2.5-VL-7B.
- Multi-dataset training (KADID-10K + SPAQ) raises VisualQuality-R1 average SRCC/PLCC from 0.777/0.814 to 0.791/0.831.
- Thurstone variant with fixed variance of one drops average SRCC to 0.770 vs 0.777 for the full sample-variance method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train VisualQuality-R1 with RL2R instead of the Q-Insight approach | NR-IQA · KADID-10K-trained, 8-dataset average | SRCC: 0.766 -> 0.777 | Helped |
| Train VisualQuality-R1 with RL2R instead of the Q-Insight approach | NR-IQA · KADID-10K-trained, 8-dataset average | PLCC: 0.803 -> 0.814 | Helped |
| Add SPAQ to KADID-10K training | VisualQuality-R1 · 8-dataset average | SRCC: 0.777 -> 0.791 | Helped |
| Add SPAQ to KADID-10K training | VisualQuality-R1 · 8-dataset average | PLCC: 0.814 -> 0.831 | Helped |
| Reduce generated responses from K=6 to K=4 (ablation) | VisualQuality-R1 · 8-dataset average | PLCC: 0.814 -> 0.811 | No clear effect |
| Reduce generated responses from K=6 to K=5 (ablation) | VisualQuality-R1 · 8-dataset average | PLCC: 0.814 -> 0.813 | No clear effect |
| Use continuous fidelity reward instead of binary reward (ablation) | VisualQuality-R1 RL2R · 8-dataset average | SRCC: 0.772 -> 0.777 | Helped |
| Use continuous fidelity reward instead of binary reward (ablation) | VisualQuality-R1 RL2R · 8-dataset average | PLCC: 0.809 -> 0.814 | Helped |
| Average quality scores rather than probabilities across individual comparisons (ablation) | VisualQuality-R1 · 8-dataset average | SRCC: 0.774 -> 0.777 | Helped |
| Average quality scores rather than probabilities across individual comparisons (ablation) | VisualQuality-R1 · 8-dataset average | PLCC: 0.805 -> 0.814 | Helped |
| Use sample variances instead of fixing variance to one (ablation) | VisualQuality-R1 Thurstone model · 8-dataset average | SRCC: 0.770 -> 0.777 | Helped |
| Use sample variances instead of fixing variance to one (ablation) | VisualQuality-R1 Thurstone model · 8-dataset average | PLCC: 0.799 -> 0.814 | Helped |
Surprising findings
- Fixing the Thurstone variance to one degrades performance on nearly all datasets, indicating sample variances from GRPO help.
- Q-Insight with rescaled MOS fails to benefit from multi-dataset training, while VisualQuality-R1 does.
Limitations
- Slow, expensive, and memory-hungry inference due to test-time reasoning
- Single fixed text prompt for all images regardless of distortion or application
- Only no-reference setting; reference-based extension left to future work
Experimental setup
Datasets: KADID-10K SPAQ
Benchmarks: BID CLIVE KonIQ-10k SPAQ Liu13 (deblurring) SRIQA-Bench Min19 (dehazing) AGIQA-3K
Metrics: SRCC PLCC
Baselines: NIQE BRISQUE UNIQUE MUSIQ MANIQA LIQE Q-Align DeQA-Score Q-Insight Qwen2.5-VL-7B
Builds on: GRPO DeepSeek-R1 Thurstone model Q-Insight Qwen2.5-VL
Closest papers by meaning
- Reasoning as Representation: Rethinking Visual Reinforcement Learning in Image Quality Assessment (ICLR 2026 · similarity 0.93)
- Q-Insight: Understanding Image Quality via Visual Reinforcement Learning (NeurIPS 2025 · similarity 0.92)
- Self-Evolving Vision-Language Models for Image Quality Assessment via Voting and Ranking (ICLR 2026 · similarity 0.91)
- Refine-IQA: Multi-Stage Reinforcement Finetuning for Perceptual Image Quality Assessment (AAAI 2026 · similarity 0.90)
- Q-Probe: Scaling Image Quality Assessment to High Resolution via Context-Aware Agentic Probing (NeurIPS 2026 · similarity 0.89)
- Adaptive Image Quality Assessment via Teaching Large Multimodal Model to Compare (NeurIPS 2024 · similarity 0.88)
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.