Erfan Esmaeili

Steering Diffusion Models via Class-Contrastive Influence for Few-Shot Medical Classification

Jeeyung Kim, Erfan Esmaeili, Qiang Qiu

When labeled data are scarce, off-the-shelf diffusion models can augment training sets for few-shot medical image classification, but not all generated samples are equally useful for the downstream task. Existing approaches largely improve synthetic data by increasing realism, di…Read more