Paper | Project Page | GitHub

Velocity Contrastive Regularization (VeCoR) is a complementary training scheme for flow-based generative modeling that augments the standard Flow Matching (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 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, particularly in low-step and lightweight settings.

On ImageNet-1K 256x256, VeCoR achieves FID=1.94 (SiT-XL/2 backbone), demonstrating significant gains in stability and image quality. It also shows consistent improvements in MS-COCO text-to-image generation.

Please refer to the official GitHub repository for environment setup, training, and sampling instructions using the provided scripts.