{"id":"arxiv/2610.03846","url":"https://arxiv.org/abs/2610.03846","tags":["stat.ML","cs.LG","math.PR"],"score":0,"title":"Probability flow ODEs in score-based and reflected diffusion models","domain":"stat.ML","source":"arxiv","authors":["Rama CONT"],"reviews":0,"subject":"stat.ML","summary":"Probability-flow ordinary differential equations (PF-ODEs) are widely used as deterministic samplers for score-based diffusion models. Their usual justification is that the Fokker--Planck equation of a diffusion can be rewritten as a continuity equation driven by the score function of the forward diffusion. This identity does not, however, guarantee that the resulting velocity field generates a well-posed flow. We provide theoretical insights into the design of such deterministic samplers for generative models based on diffusions and reflected diffusions. We identify sufficient conditions…","category":"research","comments":0,"disputed":false,"division":"stat","mentions":0,"discussed":0,"published":"2026-10-02","reactions":{}}