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)→
More from Research
- Non-Transformer Architecture Redraws ARC-AGI-1 Cost Frontier at 150M Params, Validated by Transformer Co-author — Ok_Can_1968 · 2026-08-14
- X Open-Sources For You Algorithm: Likes Worth 0.5 Points, Link Copies 20 — xiaohu · 2026-08-14
- AnyDepth: Single-Image Depth Maps with Simple Depth Transformer — tom_doerr · 2026-08-14
- SophontAI's plan: unified patient representation from multimodal medical encoders — iScienceLuvr · 2026-08-14
- 6 Essential Machine Learning Training Techniques Explained — goyalshaliniuk · 2026-08-14
- Caution Advised When Interpreting Loss Curves in Robotics — ZeYanjie · 2026-08-14