VisualQuality-R1: Reasoning-Induced Image Quality Assessment via Reinforcement Learning to Rank
Tianhe Wu, Jian Zou, Jie Liang, Lei Zhang, Kede Ma
2025-05-21
Summary
This paper talks about VisualQuality-R1, a new AI system that judges how good an image looks by thinking through its decision process, instead of just comparing it to other images.
What's the problem?
The problem is that most current models that rate image quality either need a perfect reference image to compare to, or they don't explain their reasoning in a way that matches how humans judge pictures, making their results less useful and less trustworthy.
What's the solution?
To solve this, the researchers trained VisualQuality-R1 using reinforcement learning, which means the AI learns from feedback as it tries to rank images by quality. This model doesn't need a reference image and can explain its decisions in a way that makes sense to people. It can also learn from multiple different sets of images at once.
Why it matters?
This matters because it helps create AI that can judge photos and graphics more like a human would, which is useful for everything from photo editing apps to quality control in design and media.
Abstract
VisualQuality-R1, a reasoning-induced no-reference IQA model trained via reinforcement learning, outperforms discriminative models in visual quality assessment by generating human-aligned quality descriptions and supporting multi-dataset training.