Learning Straight Flows: Variational Flow Matching for Efficient Generation

2025-11-15

AI Summary

Flow Matching (FM) struggles with high-fidelity one-step generation due to its reliance on curved trajectories, while existing improvements—such as coupling-based distribution correction and average velocity modeling—are hampered by discretization error, training instability, and poor convergence. To address this, we propose S-VFM, the first FM framework incorporating variational latent-variable modeling: a global latent code explicitly governs trajectory geometry, synergizing with average-velocity regularization to enforce end-to-end straight, consistent flows. This design inherently avoids interpolation crossing, thereby enhancing training stability and convergence. Evaluated on three major benchmarks, S-VFM consistently outperforms state-of-the-art FM and diffusion-based methods in generation quality, inference efficiency, and training robustness. By enabling interpretable and controllable one-step synthesis, S-VFM establishes a novel paradigm for deterministic, latent-guided generative modeling.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Deep Generative Models & Autoencoders
• Planning, Routing, and Scheduling: Planning with Language Models

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Federated recommendation systems and personalization

Abstract

Flow Matching has limited ability in achieving one-step generation due to its reliance on learned curved trajectories. Previous studies have attempted to address this limitation by either modifying the coupling distribution to prevent interpolant intersections or introducing consistency and mean-velocity modeling to promote straight trajectory learning. However, these approaches often suffer from discrete approximation errors, training instability, and convergence difficulties. To tackle these issues, in the present work, we propose extbf{S}traight extbf{V}ariational extbf{F}low extbf{M}atching ( extbf{S-VFM}), which integrates a variational latent code representing the ``generation overview'' into the Flow Matching framework. extbf{S-VFM} explicitly enforces trajectory straightness, ideally producing linear generation paths. The proposed method achieves competitive performance across three challenge benchmarks and demonstrates advantages in both training and inference efficiency compared with existing methods.

Problem

• Achieving one-step generation with straight trajectories
• Addressing training instability and convergence difficulties
• Improving efficiency in flow matching frameworks

Innovation

• Integrates variational latent code for generation overview
• Enforces straight trajectories for linear generation paths
• Achieves competitive performance with training efficiency