SGFlow Improves Diffusion Model Inference with Nongradient Vector Flow.

Diffusion and flow-based models are powerful for generative tasks, but their inference often requires extensive computational overhead due to the need for integration during sampling. While consistency models aim to simplify this by directly learning flow maps, existing methods face issues such as requiring model inverses or complex backpropagation through iterated model calls. Researchers introduce SGFlow, a novel method that learns flow maps without these computational burdens. SGFlow trains a model to simultaneously compute ODE solutions and implied velocities by following non-conservative dynamics, which naturally converge to the desired flow map at a stationary point. Evaluations on the CIFAR image benchmark show SGFlow achieving the best FID score at 10 sampling steps and remaining highly competitive with other state-of-the-art methods across different step counts. A key advantage of SGFlow is its unique proven stationary-point guarantee for its stopgrad-based dynamics, offering a more robust theoretical foundation.