This talk will present flow matching models, which is a specific class of image generative models, and explores the close connections between flow-based generation and denoising. We first recall how to construct a denoiser from a pretrained velocity field obtained via flow matching. We then show how this powerful denoiser can be applied to inverse problems within the Plug-and-Play (PnP) framework.
The excellent results achieved by these models in generative tasks raise the more fundamental question: what explain the good generalization capacities of generative models, i.e. their ability to generate new images, different from their training set? We will, in a second part, provide some insights into the mechanisms underlying the generalization capabilities of flow matching.