PKU-AIGIQA-4K: A Perceptual Quality Assessment Database for Both Text-to-Image and Image-to-Image AI-Generated Images

Jiquan Yuan, Jihe Li, Fanyi Yang, Xinyan Cao, Jinming Che, Jinlong Lin, Xixin Cao; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 3362-3371

AIGIQA, aimed at assessing the perceptual quality of AI-generated images (AIGIs), is essential for image quality screening and the evaluation of generative models. However, there are two issues in exsiting work: 1) existing AIGIQA databases are limited to images generated by text-to-image (T2I) generative models, lacking exploration of the evaluation of images generated by image-to-image (I2I) generative models; 2) existing image quality assessment (IQA) methods often fail to fully utilize the information provided by image prompts when simultaneously assessing T2I and I2I AIGIs. To address these problems, we first establish a large scale perceptual quality assessment database for both T2I and I2I AIGIs, named AIGIQA-4K. Second, we propose a novel partial-reference IQA (PR-IQA) method in this paper. Finally, leveraging the AIGIQA-4K database, we conduct extensive benchmark experiments using several pretrained models and the current IQA methods to explore the applicability of these methods for evaluating the quality of both the T2I and I2I AIGIs. The PKU-AIGIQA-4K database and codes are released on https://github.com/jiquan123/AIGIQA4K.