Diffusion vs flow matching trained from scratch shows big difference

View on Reddit

Trained both diffusion and flow matching generative models from scratch on COCO-2017 dataset using CLIP ViT-L and FLUX VAE. Reports significant quality differences between the two approaches.