Joint Diffusion Sampling via Positive-Unlabeled Guidance for Multi-Modal Data

ICML 2025 Workshop FM4LS Submission44 Authors

Keywords: joint modeling, joined models, weakly-supervised learning

TL;DR: We use a positive-unlabeled classifier to coordinate pretrained diffusion models for images, molecules, and Boolean values to generate uni- and multi-modal data.

Abstract: Multi-modal generative models typically require abundant training data from multi-modal joint distributions, which is often unavailable in the life sciences. We propose to treat each modality as a marginal distribution and correct their independent diffusion processes to sample from their joint distribution. Specifically, we introduce "joint diffusion sampling," a method that generates a sample from joint distributions using pre-trained models for individual (uni-modal) marginal distributions and minimal data from the (multi-modal) joint distribution. We demonstrate preliminary uni- and multi-modal results for images, molecules, and Boolean values, and discuss multi-modal applications of our approach.

Submission Number: 44