An Evaluation of Text-Guided Image Editing Models

Text-guided image editing, which modifies images based solely on an input image and a target text prompt, will soon become a crucial function in computer vision and image processing. However, evaluations of existing models are still very limited. In this study, we systematically evaluate state-of-the-art models for text-guided image editing through both subjective and objective quality assessments, marking the first comprehensive evaluation to address all four essential criteria: Alignment, Preservation, Perception, and Acceptability, with a robust participant base for reliability. This methodology enables an in-depth comparison of models, showing that Imagic and Forgedit consistently lead in all criteria. Furthermore, our objective analysis, utilizing a wide range of Image Quality Assessment (IQA) and alignment metrics, reveals significant limitations in the existing metrics' abilities to evaluate accurately the performance of text-guided image editing models.