Browse Teach Curate

Anti-Editing Concept Erasure in Text-to-Image Diffusion Models

Extending concept erasure so an unlearned concept is blocked not only at prompt-driven generation but also under image editing: ACE injects erasure guidance into both conditional and unconditional noise predictions of a fine-tuned Stable Diffusion model, with a prior constraint loss and stochastic prior-guided correction guidance to prevent erosion of unrelated concepts.

Optimus Prime Text

ACE: Anti-Editing Concept Erasure in Text-to-Image Models Zihao Wang1 Yuxiang Wei1 Fan Li2 Renjing

Recent advance in text-to-image diffusion models have significantly facilitated the generation of high-quality images, but also raising concerns about the illegal creation of harmful content, such as…