The paper introduces AIGIQA-20K, a large-scale database for assessing the quality of AI-generated images (AIGIs). The database includes 20,000 AIGIs and 420,000 subjective ratings from 21 subjects, covering 15 popular generative models. The authors dynamically adjust hyper-parameters such as classifier-free guidance (CFG), iteration epochs, and output image resolution to comprehensively evaluate the visual distortion of AIGIs. They also conduct benchmark experiments to assess the performance of 16 mainstream AIGI quality models against human perception. The study highlights the importance of considering both perceptual quality and text-to-image alignment in AIGI quality assessment. The database and experiments aim to inspire robust quality indicators for AIGIs and advance the field of AIGC in vision.The paper introduces AIGIQA-20K, a large-scale database for assessing the quality of AI-generated images (AIGIs). The database includes 20,000 AIGIs and 420,000 subjective ratings from 21 subjects, covering 15 popular generative models. The authors dynamically adjust hyper-parameters such as classifier-free guidance (CFG), iteration epochs, and output image resolution to comprehensively evaluate the visual distortion of AIGIs. They also conduct benchmark experiments to assess the performance of 16 mainstream AIGI quality models against human perception. The study highlights the importance of considering both perceptual quality and text-to-image alignment in AIGI quality assessment. The database and experiments aim to inspire robust quality indicators for AIGIs and advance the field of AIGC in vision.