SGFlow: Efficient Flow Map Learning Without Gradient Backpropagation

kwangmoo_yi · x · 2026-08-06

Diffusion and flow-based models incur significant inference overhead due to sampling integration. Consistency models address this by learning flow maps, but existing methods face computational challenges like model inverses or backpropagation through iterated calls.

The ICLR 2026 paper Flow Map Learning via Nongradient Vector Flow introduces SGFlow. It bypasses explicit invertibility constraints and expensive differentiation by training a model from scratch to compute ODE solutions and implied velocity via non-conservative dynamics with a stationary point. On the CIFAR benchmark, SGFlow achieves the best FID at 10 sampling steps and remains competitive across other step counts, uniquely offering a proven stationary-point guarantee for its stopgrad-based dynamics.

Related event: ICLR Paper Introduces SGFlow for Efficient Flow Matching Without Backpropagation(2 posts)→

Original post →

More from Research

Research channel →