Expanding Flow Maps: Unifying Fast Generation Across Continuous and Discrete State Spaces

arXiv 2607.21585·medium signal

Sophia Tang and Pranam Chatterjee present Expanding Flow Maps, generalizing flow-map generative modeling beyond the parameterizations that existing methods are restricted to, covering both continuous and discrete state spaces. Flow maps are the mechanism behind few-step generation — collapsing a many-step sampling trajectory into a learned direct jump. The discrete-space coverage is the notable part, since most fast-sampling work has been confined to continuous domains like images and audio.