Diffusion-Derep: Removing Sample Replication During Diffusion Inference

Recent advancements in research on diffusion-based generative models have yielded excellent results in image, video, and 3D generation. Several open-source models for real-world images have recently been released, generating photorealistic results. However, in the medical domain, most major foundation diffusion models are held privately due to the data replication issue inherent in diffusion models. The data replication issue refers to a sampling problem observed in diffusion models, where the model tends to generate samples closely resembling the training data. This phenomenon becomes pronounced when a diffusion model is trained on a small dataset containing only a few hundred images. In this work, we address the data replication issue in diffusion models by proposing a novel inference process guidance algorithm. Specifically, during the inference process, we steer the generated samples away from the training samples by utilizing a similarity loss that quantifies the distance between the generated sample and the embedded representations of the training examples. To the best of our knowledge, this is the first attempt to tackle the sample replication problem in diffusion models. Our method can be easily integrated with existing methods, enhancing their performance in generating new, novel samples. We present results across 8 datasets across different domains and show that utilizing our sampling algorithm significantly reduces data mitigation and observed in our FID and SSCD experiments. Code: https://github.com/Vibashan/MedDataset