2 papers
cs.LG2025
Gradient Variance Reveals Failure Modes in Flow-Based Generative Models
Teodora Reu, Sixtine Dromigny, Michael Bronstein +1
Rectified Flows learn ODE vector fields whose trajectories are straight between source and target distributions, enabling near one-step inference. We show that this straight-path o…
cs.LG2024
Metric Flow Matching for Smooth Interpolations on the Data Manifold
Kacper KapuÅniak, Peter Potaptchik, Teodora Reu +5
Matching objectives underpin the success of modern generative models and rely on constructing conditional paths that transform a source distribution into a target distribution. Des…