Imagic is a novel method for text-based real image editing that can perform complex, non-rigid semantic edits on a single high-resolution input image. Unlike previous methods that are limited to specific editing types, synthetically generated images, or require multiple input images, Imagic can change the posture, composition, and appearance of objects within an image while preserving its original characteristics. The method leverages a pre-trained text-to-image diffusion model to generate a text embedding that aligns with both the input image and the target text. This embedding is then fine-tuned to better reconstruct the input image and finally interpolated between the target text embedding and the optimized embedding to produce the final edited image. Imagic demonstrates high-quality and versatile editing capabilities on various domains and introduces TEDBench, a challenging benchmark for evaluating text-based image editing methods. User studies show that human raters prefer Imagic over other leading methods on TEDBench.Imagic is a novel method for text-based real image editing that can perform complex, non-rigid semantic edits on a single high-resolution input image. Unlike previous methods that are limited to specific editing types, synthetically generated images, or require multiple input images, Imagic can change the posture, composition, and appearance of objects within an image while preserving its original characteristics. The method leverages a pre-trained text-to-image diffusion model to generate a text embedding that aligns with both the input image and the target text. This embedding is then fine-tuned to better reconstruct the input image and finally interpolated between the target text embedding and the optimized embedding to produce the final edited image. Imagic demonstrates high-quality and versatile editing capabilities on various domains and introduces TEDBench, a challenging benchmark for evaluating text-based image editing methods. User studies show that human raters prefer Imagic over other leading methods on TEDBench.