The authors studied the mechanism of diffusion transformer (DiT) editors such as Flux Kontext, which use double-stream blocks that mix text and image tokens.
Forcing image editors to copy the reference prevents harmful edits
The authors show that maximizing target-to-reference attention in early transformer blocks causes diffusion transformers to reproduce the input rather than execute edits.
Big Tech
Salma Abdel Magid · Sanghyuk Chun · Eric Slyman · Kushal Kafle · Olga Russakovsky
Princeton University · Adobe Research
Research Digest··2 min read
The authors propose Copying-as-Defense (CaD), which adds imperceptible perturbations to images that force diffusion transformer editors to copy the reference instead of editing it.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.