Diffusion-Based Image Editing

The field studies how diffusion models can make targeted changes to existing images, from replacing objects and editing text to transferring style, adjusting faces, and relighting scenes. Approaches include instruction-tuned editors, reference and mask conditioning, and training-free techniques that steer attention, latent representations, or denoising trajectories. Recent work also uses reinforcement learning, multi-view geometry, and faster few-step models to handle complex edits, preserve scene structure, and support high-resolution or interactive use. A central challenge is balancing faithful instruction following with preservation of identity and unedited regions, while keeping inference efficient and evaluation reliable.