Training One‑Step Generative Models Without CFG, DMD, GAN, Drifting, or MeanFlow

The paper introduces TBSM, a lightweight direction‑tracking network that learns fake‑to‑real movement fields from three‑sample scattering events, enabling single‑forward‑pass generation and achieving state‑of‑the‑art FID scores on ImageNet and a 20B text‑to‑image model without relying on CFG, DMD, GAN, drifting, or mean‑flow techniques.

PixelDiTQwen-ImageTBSM

0 likes · 12 min read

Training One‑Step Generative Models Without CFG, DMD, GAN, Drifting, or MeanFlow

Machine Heart

Aug 9, 2026 · Artificial Intelligence

How to Train a One‑Step Generative Model Without CFG, DMD, GAN, or Drifting

The paper introduces TBSM, a lightweight direction‑tracking network that learns per‑sample Fake‑to‑Real vectors to guide a generator, enabling single‑forward (NFE=1) image synthesis with FID 1.92 on ImageNet‑512 and high‑quality text‑to‑image results, all without CFG, DMD, GAN, or drifting methods.

TBSMdiffusion modelsdirection tracking

0 likes · 11 min read

How to Train a One‑Step Generative Model Without CFG, DMD, GAN, or Drifting