Reproduce “The Generation Phases of Flow Matching: a Denoising Perspective” on OpenResearch

About this paper

Anne Gagneux, Ségolène Martin, Rémi Gribonval, Mathurin Massias

Researchers from Inria, ENS de Lyon, CNRS, Université Claude Bernard Lyon 1, LIP, and Technische Universität Berlin empirically investigate how flow matching generative models achieve high-quality results by analyzing their generation process through a denoising lens. The study reveals that the effectiveness of these models hinges on distinct temporal denoising phases, with the intermediate phase being crucial for learned models to maintain expressive velocity fields and achieve high-fidelity generation.

Key insights

• While theoretically equivalent, empirical performance in flow matching is highly dependent on the choice of denoising loss and network parametrization, with the standard FM loss and a residual parametrization proving most effective.
• The generation process has distinct temporal phases: early stages establish global structure via low-frequency changes, while late stages refine details via high-frequency adjustments, and the intermediate phase is crucial for learned models.
• Learned flow matching models do not replicate the sharp early peak in velocity field regularity (Jacobian spectral norm) seen in theoretical closed-form solutions; instead, they maintain higher regularity in the intermediate regime, which is vital for high-quality generation.

Summary by alphaXiv.