Deep generative models have achieved remarkable success over the last few years. Among recent techniques, Flow Matching has emerged as an effective framework with a simple, simulation-free training objective. It learns how to map samples from a simple base distribution to a complex target distribution by modeling the velocity field of an interpolating probability path. New samples are then generated by numerically integrating an ordinary differential equation. This thesis studies Flow Matching (FM) and develops two related directions. First, it presents Acceleration Matching (AM), a second-order extension designed for trajectory inference from multiple unpaired distributions observed at different times. By lifting the dynamics to phase space, AM learns an acceleration field whose associated stochastic process reproduces the prescribed positional marginals. After a detailed mathematical derivation, the method is tested on several benchmark problems, showing that it is competitive with other common trajectory-inference approaches. Second, the thesis investigates image super-resolution as a preliminary step towards applying these methods to cosmological simulation data. For this purpose, we apply Conditional Flow Matching (CFM), a variant of FM where the velocity field is conditioned on the low-resolution image. We theoretically motivate how this conditioning preserves the dependence between paired low- and high-resolution images, and evaluate the method qualitatively on MNIST, CIFAR-10, and DIV2K. Overall, the thesis illustrates how learned continuous-time dynamics can be adapted both to multi-marginal trajectory inference and to conditional image generation.

Learning Flows of Probability Measures: from Super-Resolution to Acceleration Matching

DAZZINI, BARTOLO
2025/2026

Abstract

Deep generative models have achieved remarkable success over the last few years. Among recent techniques, Flow Matching has emerged as an effective framework with a simple, simulation-free training objective. It learns how to map samples from a simple base distribution to a complex target distribution by modeling the velocity field of an interpolating probability path. New samples are then generated by numerically integrating an ordinary differential equation. This thesis studies Flow Matching (FM) and develops two related directions. First, it presents Acceleration Matching (AM), a second-order extension designed for trajectory inference from multiple unpaired distributions observed at different times. By lifting the dynamics to phase space, AM learns an acceleration field whose associated stochastic process reproduces the prescribed positional marginals. After a detailed mathematical derivation, the method is tested on several benchmark problems, showing that it is competitive with other common trajectory-inference approaches. Second, the thesis investigates image super-resolution as a preliminary step towards applying these methods to cosmological simulation data. For this purpose, we apply Conditional Flow Matching (CFM), a variant of FM where the velocity field is conditioned on the low-resolution image. We theoretically motivate how this conditioning preserves the dependence between paired low- and high-resolution images, and evaluate the method qualitatively on MNIST, CIFAR-10, and DIV2K. Overall, the thesis illustrates how learned continuous-time dynamics can be adapted both to multi-marginal trajectory inference and to conditional image generation.
2025
Learning Flows of Probability Measures: from Super-Resolution to Acceleration Matching
Generative Modeling
Flow Matching
Deep Learning
File in questo prodotto:
File Dimensione Formato  
DAZZINI_BARTOLO.pdf

accesso aperto

Dimensione 14.93 MB
Formato Adobe PDF
14.93 MB Adobe PDF Visualizza/Apri

The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.12608/115831