How Q-Insight Uses Reinforcement Learning to Make AI Truly Understand Image Quality

Background

Generative AI and multimodal large models have made it possible to synthesize high‑quality images, but evaluating the visual quality of these machine‑generated images remains difficult. Traditional image‑quality assessment (IQA) methods either output a single scalar score without explanation or rely on massive text‑description datasets, which limits interpretability and generalisation.

Problem

In the audio‑video processing pipeline every stage—from capture to playback—depends on human visual perception. Existing IQA approaches cannot reason about the underlying causes of visual degradation and often fail on out‑of‑domain data.

Q‑Insight: A Multimodal Large‑Model IQA Solution

Q‑Insight, proposed by researchers from Peking University and Volcano Engine’s Multimedia Lab, treats the quality score as a guiding signal that drives the model to *think* about the root causes of image quality. The model is built on a multimodal large language model (LLM) and is trained with reinforcement learning.

Key Innovations

First integration of reinforcement learning into image‑quality assessment.

Introduction of the Group Relative Policy Optimization (GRPO) algorithm, which learns relative quality preferences from groups of images and eliminates the need for large‑scale text supervision.

Multi‑output design: the model produces a quality score, identifies degradation type, performs pairwise comparison, and generates a step‑by‑step reasoning chain.

Methodology

During training the model receives a *group* of images (e.g., three to five samples). For each image it predicts a quality score and a textual explanation. The reward signal is derived from human‑perceived quality (e.g., MOS) and from the correctness of the degradation classification. GRPO updates the policy by comparing the predicted relative ordering of the group with the ground‑truth ordering, using a policy‑gradient loss that encourages higher‑quality images to receive higher scores. Because the reward is based on relative preferences, the model does not require explicit absolute labels for every image, reducing dependence on costly annotation.

Training Details

Q‑Insight is fine‑tuned on publicly available IQA datasets (such as LIVE, TID2013, KonIQ‑10k) without any additional text‑description supervision. The reinforcement‑learning loop runs for 50 k steps with a learning rate of 1e‑5, and the GRPO optimizer uses a clipping parameter of 0.2. The model’s multimodal backbone remains frozen except for the quality‑assessment heads, which keeps the number of trainable parameters under 10 M.

Experimental Evaluation

Extensive experiments on several public benchmarks demonstrate that Q‑Insight:

Achieves higher Pearson and Spearman correlation with human MOS than state‑of‑the‑art NR‑IQA methods, especially on out‑of‑domain test sets.

Detects degradation types (e.g., Gaussian noise, JPEG compression) with >90 % accuracy, outperforming specialised degradation classifiers.

Performs zero‑shot pairwise image comparison, providing detailed quality reasoning without any additional fine‑tuning.

All quantitative results are reported in the paper https://arxiv.org/pdf/2503.22679 .

Implications for Video‑Cloud Services

By embedding Q‑Insight into a video‑cloud stack, downstream components such as generative quality‑enhancement models, immersive audio models, and intelligent video encoders can receive fine‑grained quality feedback. This enables adaptive encoding, automated quality‑enhancement triggers, and more reliable user‑experience monitoring while reducing the need for manual quality‑assessment pipelines.

Conclusion

Q‑Insight shows that reinforcement learning combined with the GRPO algorithm can endow multimodal large models with genuine visual‑quality reasoning, delivering accurate scores, degradation diagnostics, and interpretable explanations without large text‑annotation corpora.

Code example

收
藏
，
分
享
、
在
看
，
给
个
三
连
击呗！