Efficient Flow Matching using Latent Variables

The paper presents Latent-CFM, an efficient flow matching model that leverages features from pretrained deep latent variable models to improve image generation tasks. By conditioning on these latent features, Latent-CFM achieves better generation quality with reduced training time and computational costs compared to existing flow matching models. This approach not only enhances generative modeling for natural images but also applies to physical processes, as demonstrated with a 2D Darcy flow dataset, offering interpretability in conditional image generation.