Q-Hawkeye: Reliable Visual Policy Optimization for Image Quality Assessment

TL;DR

Q-Hawkeye tackles reliability and perceptual grounding in RL-based image quality assessment by introducing Uncertainty-Aware Dynamic Optimization and Perception-Aware Optimization. It uses rollout score variance to adaptively reweight updates and constructs original–degraded image pairs with an implicit perception loss to ground judgments in visual evidence, all within a GRPO-based framework. The approach yields superior, robust IQA performance across eight benchmarks while maintaining data efficiency and cross-dataset generalization, with modest computational overhead. This work advances practical IQA for real-world vision tasks and AI-generated content quality control by improving both stability and visual grounding of perceptual scores.

Abstract

Image Quality Assessment (IQA) predicts perceptual quality scores consistent with human judgments. Recent RL-based IQA methods built on MLLMs focus on generating visual quality descriptions and scores, ignoring two key reliability limitations: (i) although the model's prediction stability varies significantly across training samples, existing GRPO-based methods apply uniform advantage weighting, thereby amplifying noisy signals from unstable samples in gradient updates; (ii) most works emphasize text-grounded reasoning over images while overlooking the model's visual perception ability of image content. In this paper, we propose Q-Hawkeye, an RL-based reliable visual policy optimization framework that redesigns the learning signal through unified Uncertainty-Aware Dynamic Optimization and Perception-Aware Optimization. Q-Hawkeye estimates predictive uncertainty using the variance of predicted scores across multiple rollouts and leverages this uncertainty to reweight each sample's update strength, stabilizing policy optimization. To strengthen perceptual reliability, we construct paired inputs of degraded images and their original images and introduce an Implicit Perception Loss that constrains the model to ground its quality judgments in genuine visual evidence. Extensive experiments demonstrate that Q-Hawkeye outperforms state-of-the-art methods and generalizes better across multiple datasets. The code and models will be made available.

Figures (15)

• Figure 1: Illustration of predictive uncertainty during training. Sample A shows a high-variance score distribution with inconsistent reasoning, while Sample B shows a low-variance distribution with consistent reasoning.
• Figure 2: Visual perception analysis of existing state-of-the art methods and our Q-Hawkeye on the KonIQ dataset. For each method, we sample responses conditioned on the original images, and compute the mean log-probability of the same responses under the original and degraded conditions. Histograms show the distributions and the mean gap between them.
• Figure 3: Overview of the proposed Q-Hawkeye framework. For each image--prompt pair, the policy model produces $K$ reasoning trajectories and quality scores. The Uncertainty-Aware Dynamic Optimization module computes the score variance as an uncertainty signal and uses it to rescale sample-wise advantages. In parallel, Perception-Aware Optimization constructs original--degraded image pairs and maximizes the KL divergence between their output distributions, enforcing sensitivity to visual degradations.
• Figure 4: Training dynamics of uncertainty-aware optimization.Average reward (left) and reward std (right) for vanilla GRPO, our uncertainty-aware weighting, and a reverse weighting baseline.
• Figure 5: PLCC performance on each dataset when varying the number of generated responses $K$.