Main content start

To diffuse or not? One-step generative modeling via Wasserstein gradient flows

Date
Tue October 20th 2026, 4:00pm
Location
CoDa E160
Speaker
Emmanuel Candès, Stanford Math and Statistics

Diffusion models and flow-based methods have demonstrated impressive generative capabilities, especially for images, but sampling remains expensive because it requires many iterative updates. This talk introduces W-Flow, a framework for training a generator that maps samples from a simple reference distribution to samples from a target data distribution in a single step. W-Flow first evolves the reference distribution toward the target through a Wasserstein gradient flow that minimizes an energy functional, then trains a static neural generator to compress this evolution into one-step generation. We use the Sinkhorn divergence as the energy functional, yielding an efficient optimal transport-based update rule that captures global distributional discrepancies and improves coverage of the target distribution. Empirically, W-Flow achieves a new state of the art for one-step ImageNet 256-by-256 generation, with an FID of 1.29, alongside improved mode coverage and domain transfer. Compared with multi-step diffusion models achieving similar FID scores, W-Flow samples approximately 100 times faster.

This is joint work with Jiaqi Han, Puheng Li, Qiushan Guo, Renyuan Xu, and Stefano Ermon.