VeCoR - Velocity Contrastive Regularization for Flow Matching

2025-11-24

AI Summary

Flow matching (FM) suffers from error accumulation along integration trajectories during velocity field learning, causing generated samples to deviate from the data manifold—especially under low-step sampling or with lightweight models, leading to substantial degradation in sample quality. To address this, we propose a bidirectional attraction-repulsion training paradigm, introducing Velocity Contrastive Regularization (VeCoR), which imposes dual supervision on the velocity field: alignment of positive velocity pairs and repulsion of negative ones. This upgrades the conventional unidirectional attraction objective to a geometrically grounded bidirectional constraint, thereby stabilizing trajectory evolution and enhancing manifold consistency. Experiments on text-to-image generation over ImageNet-1K and MS-COCO demonstrate relative FID improvements of 22–35% (ImageNet) and 32% (COCO) over FM baselines, alongside accelerated convergence and improved training stability. Our key contribution is the first integration of contrastive learning into FM-based velocity field optimization, significantly boosting perceptual fidelity and generalization under low computational overhead.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Learning with Manifolds
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Flow Matching (FM) has recently emerged as a principled and efficient alternative to diffusion models. Standard FM encourages the learned velocity field to follow a target direction; however, it may accumulate errors along the trajectory and drive samples off the data manifold, leading to perceptual degradation, especially in lightweight or low-step configurations. To enhance stability and generalization, we extend FM into a balanced attract-repel scheme that provides explicit guidance on both "where to go" and "where not to go." To be formal, we propose extbf{Velocity Contrastive Regularization (VeCoR)}, a complementary training scheme for flow-based generative modeling that augments the standard FM objective with contrastive, two-sided supervision. VeCoR not only aligns the predicted velocity with a stable reference direction (positive supervision) but also pushes it away from inconsistent, off-manifold directions (negative supervision). This contrastive formulation transforms FM from a purely attractive, one-sided objective into a two-sided training signal, regularizing trajectory evolution and improving perceptual fidelity across datasets and backbones. On ImageNet-1K 256$ imes$256, VeCoR yields 22% and 35% relative FID reductions on SiT-XL/2 and REPA-SiT-XL/2 backbones, respectively, and achieves further FID gains (32% relative) on MS-COCO text-to-image generation, demonstrating consistent improvements in stability, convergence, and image quality, particularly in low-step and lightweight settings. Project page: https://p458732.github.io/VeCoR_Project_Page/

Problem

• Improves flow matching stability by preventing trajectory deviation errors
• Enhances perceptual fidelity through contrastive velocity regularization
• Addresses performance degradation in lightweight and low-step configurations

Innovation

• Velocity Contrastive Regularization enhances Flow Matching
• Two-sided supervision attracts positive and repels negative directions
• Improves stability and image quality in lightweight configurations