DragonDiffusion

DragonDiffusion is a refined image editing solution based on the diffusion model, which supports multiple editing modes such as object movement, object resizing, object appearance replacement, and content dragging. The editing signal is transformed into gradients via a feature correspondence loss, modifying the intermediate representation of the diffusion model. The feature correspondence loss considers multiple scales of semantic and geometric alignment, and adds a cross-branch self-attention mechanism to maintain consistency between the original image and the edited results.

DragonDiffusion can be widely used in real-image editing, including object movement, object resizing, object appearance replacement, and content dragging.