TBSM: A New Training Method for One-Step Generative Models Without Complex Distillation

This article introduces a new method called TBSM (Three-Body Scattering for Generative Modeling) for training one-step generative models. It trains a lightweight tracker online to learn the fake-to-real transfer field, enabling high-quality image generation in a single forward pass. On ImageNet-256, TBSM achieves FID 1.63 (latent) and 2.23 (pixel) with NFE=1, demonstrating its efficiency and effectiveness.