RMFlow: Refined Mean Flow by a Noise-Injection Step for Multimodal Generation

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Huang, Yuhao, Wang, Shih-Hsin, Bertozzi, Andrea L., Wang, Bao
Natura: Preprint
Pubblicazione: 2026
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866912866664185856
author Huang, Yuhao
Wang, Shih-Hsin
Bertozzi, Andrea L.
Wang, Bao
author_facet Huang, Yuhao
Wang, Shih-Hsin
Bertozzi, Andrea L.
Wang, Bao
contents Mean flow (MeanFlow) enables efficient, high-fidelity image generation, yet its single-function evaluation (1-NFE) generation often cannot yield compelling results. We address this issue by introducing RMFlow, an efficient multimodal generative model that integrates a coarse 1-NFE MeanFlow transport with a subsequent tailored noise-injection refinement step. RMFlow approximates the average velocity of the flow path using a neural network trained with a new loss function that balances minimizing the Wasserstein distance between probability paths and maximizing sample likelihood. RMFlow achieves near state-of-the-art results on text-to-image, context-to-molecule, and time-series generation using only 1-NFE, at a computational cost comparable to the baseline MeanFlows.
format Preprint
id arxiv_https___arxiv_org_abs_2602_00849
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle RMFlow: Refined Mean Flow by a Noise-Injection Step for Multimodal Generation
Huang, Yuhao
Wang, Shih-Hsin
Bertozzi, Andrea L.
Wang, Bao
Machine Learning
Artificial Intelligence
Numerical Analysis
68Txx
Mean flow (MeanFlow) enables efficient, high-fidelity image generation, yet its single-function evaluation (1-NFE) generation often cannot yield compelling results. We address this issue by introducing RMFlow, an efficient multimodal generative model that integrates a coarse 1-NFE MeanFlow transport with a subsequent tailored noise-injection refinement step. RMFlow approximates the average velocity of the flow path using a neural network trained with a new loss function that balances minimizing the Wasserstein distance between probability paths and maximizing sample likelihood. RMFlow achieves near state-of-the-art results on text-to-image, context-to-molecule, and time-series generation using only 1-NFE, at a computational cost comparable to the baseline MeanFlows.
title RMFlow: Refined Mean Flow by a Noise-Injection Step for Multimodal Generation
topic Machine Learning
Artificial Intelligence
Numerical Analysis
68Txx
url https://arxiv.org/abs/2602.00849