Conceptual

Diffusion Models for Unbalanced Multimodal Paired Scientific Data Generation

UB-Diff, a diffusion-based generative model that produces paired multi-modal scientific data (for example the spatial velocity map and the seismic waveform in subsurface imaging) even when the training data are unbalanced across modalities. Its key idea is a one-in-two-out encoder-decoder that maps available data into a shared co-latent representation guaranteeing the two output modalities stay paired, after which the diffusion process samples new pairs from that co-latent space; it is evaluated on the OpenFWI dataset and improves Frechet Inception Distance and pairwise-consistency over prior multi-modal generators.