W-Flow: One-Step Generative Modeling via Wasserstein Gradient Flows Achieves SOTA ImageNet FID of 1.29

Stanford researchers introduced W-Flow, a framework that achieves one-step image generation with state-of-the-art ImageNet 256x256 FID of 1.29, making sampling approximately 100x faster than multi-step diffusion models. W-Flow trains a generator to transform reference distribution samples into target distribution samples in a single step using Wasserstein gradient flows. The energy functional is instantiated with Sinkhorn divergence for efficient optimal-transport-based updates. The work includes formal proofs that finite-sample training dynamics converge to continuous-time distributional dynamics. Published on arXiv 2605.11755 on 2026-05-13.