Enabling Instructional Image Editing with In-Context Generation in Large Scale Diffusion Transformer

Change the background to Hawaii scenery. Dress in Aloha Shirt, Hawaiian shorts and surf on board. 2 1 3 4 5 Replace the boy with SpongeBob and make it a comic book photo. Add the text Aloha Hawaii on the bottom in bold white color. Holding a cup of tea, eye closed. Wears a diamond earring, and a golden ruby crown. Make her hair dark green and her clothes checked. Girl is on the beach, colorful cloud in sky. What if it looks like watercolor painting? ID Consistent Editing Multi-turn Editing Multi-task Editing

Figure 1: We introduce ICEdit, a novel method that achieves state-of-the-art instruction-based image editing with only 0.1% training data required by previous SOTA methods, demonstrating exceptional generalization. The first row illustrates a series of multi-turn edits, executed with high precision, while the second and third rows highlight diverse, visually impressive editing results from our method.