Equilibrium Matching simplifies flow matching and hits FID 1.96 on ImageNet 256

du_yilun · x · 2025-10-06

Researchers propose Equilibrium Matching (EqM), a new generative method that is simpler than flow matching while outperforming it, achieving FID 1.96 on ImageNet 256×256.

The key idea is to learn a single static EBM (energy-based model) landscape for generation, which enables a simple gradient-based sampling procedure at inference time.

Related event: Equilibrium Matching Surpasses Flow Matching with FID 1.96(2 posts)→

Original post →

More from Research

Research channel →