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

2026-01-31

AI Summary

This work addresses the challenge that MeanFlow struggles to balance generation quality and diversity in multimodal settings under a single function evaluation (1-NFE). To overcome this limitation, the authors propose a novel 1-NFE generative framework that first models a coarse-grained transport via a neural network–parameterized mean velocity field and then refines outputs through a tailored noise injection mechanism to enhance sample fidelity. A new loss function is introduced that simultaneously minimizes the Wasserstein distance between probability paths and maximizes sample likelihood. The method achieves near state-of-the-art performance across diverse tasks—including text-to-image synthesis, context-to-molecule generation, and time-series modeling—using only one NFE, with computational overhead comparable to the original MeanFlow.

Technology Category

• Natural Language Processing: Generation
• Machine Learning: Multimodal Learning
• Computer Vision: Generative Adversarial Networks (GANs) for Vision

Application Category

• Social Networks and Social Media: Generative AI / large language models and their impact on social systems
• Web Mining and Content Analysis: Web data generation and simulation
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG

Abstract

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.

Problem

• MeanFlow
• multimodal generation
• 1-NFE
• image generation
• generative modeling

Innovation

• RMFlow
• MeanFlow
• noise-injection refinement
• multimodal generation
• Wasserstein distance