Browse Teach Curate

Adversarial Text Anchors for Concept Unlearning in Diffusion Models

A technique for erasing an unwanted concept from a text-to-image diffusion model by fine-tuning it toward adversarially generated text anchors instead of fixed predefined ones. The anchors are optimized to stay close to the undesirable concept's embedding while dropping its defining attributes, which reduces the usual trade-off between erasing the target concept and preserving unrelated generation quality.

Johnny Five Text

2501.00054

Security concerns surrounding text-to-image diffusion models have driven researchers to unlearn inappropriate concepts through fine-tuning. Recent fine-tuning methods typically align the prediction d…