Tag
The paper introduces a method for controlled outlier generation using diffusion models by manipulating likelihood through Radon-Nikodym derivatives, enabling the creation of low-likelihood samples without retraining.