Residual-space adaptation cuts forgetting in continual diffusion personalization
Restricting updates to layers identified by residual dynamics keeps performance while cutting trainable parameters by up to 72%.
Stories per day · 30 days
1 stories, newest first · one card per story