Tag
This paper introduces SGFlow, a method for learning flow maps for diffusion models that avoids invertibility constraints and backpropagation through model iterations, achieving competitive FID scores on CIFAR with a proven stationary-point guarantee.