Training-Free Image Editing: Stable Flow Revolutionizes the Field
Stable Flow is a training-free image editing method leveraging the Diffusion Transformer (DiT) model. It achieves various image editing operations, including non-rigid editing, object addition, removal, and global scene editing, by selectively injecting attention features. Unlike UNet-based models, DiT lacks a coarse-to-fine synthesis structure. The researchers propose an automatic method to identify "vital layers" crucial for image formation within DiT. By injecting features from the source image's generative trajectory into the edited image's trajectory, Stable Flow enables consistent and stable edits. Furthermore, it introduces an improved image inversion method for real-image editing. Experiments demonstrate Stable Flow's effectiveness across diverse applications.
Read more