AIGC Image Quality Assessment via Image-Prompt Correspondence

Fei Peng, Huiyuan Fu, Anlong Ming, Chuanming Wang, Huadong Ma, Shuai He, Zifei Dou, Shu Chen; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 6432-6441

In the rapidly evolving landscape of deep learning generative models such as Generative Adversarial Networks (GANs) and diffusion models have significantly advanced the capabilities of Artificial Intelligence Generated Content (AIGC). These technologies have streamlined the creative process enabling AI to autonomously produce a diverse range of content with minimal human input. Despite the remarkable progress in AI-generated images (AIGIs) evaluating the quality of AIGIs remains a complex challenge. Traditional image quality assessment (IQA) focusing on aspects like distortion and blurriness are insufficient for capturing the correspondence between AIGIs and their prompts. To address this we propose a novel AIGC image quality assessment (AIGCIQA) framework that emphasizes the correspondence between images and prompts. Utilizing the CLIP model's pre-trained image and text encoders our method effectively measures the correspondence between visual and textual inputs. By transforming the assessment into classification probabilities and subsequently into a precise regression task our method enhances the CLIP model's performance in AIGCIQA. Our method's effectiveness is confirmed by its first place in the image track of the NTIRE 2024 Quality Assessment for AI-Generated Content challenge and its state-of-the-art (SOTA) performance on benchmark datasets AGIQA-1K AGIQA-3K and AIGCIQA2023. This research represents a significant advancement in the field offering an efficient and versatile tool for the evaluation of AIGIs and contributing to the ongoing development of AIGC technologies. Our codes are available at https://github.com/pf0607/IPCE.