Latent Flow

Latent flow models are generative models that learn complex data distributions by transforming a simple latent space into a high-dimensional data space using learned flow fields. Current research focuses on improving efficiency and scalability through techniques like flow matching in latent spaces, incorporating Gaussian processes for improved sample quality, and leveraging neural ordinary/stochastic differential equations for modeling temporal dynamics. These advancements are impacting diverse fields, enabling improved forecasting in medical imaging, efficient exploration of chemical space for drug discovery, and enhanced representation learning for various data types including images, videos, and time series.