Three-Body Scattering for Generative Modeling

Researchers introduced Three-Body Scattering Modeling (TBSM), a new one-step generative training framework.

TBSM learns a transport field that guides synthetic samples toward real data, instead of relying on adversarial training, fixed denoising paths, or teacher distillation.

The approach delivers strong results on ImageNet-256 and large text-to-image models, including PixelDiT-XL and DiT-XL.

It offers a more stable and scalable path for high-quality image and text-to-image generation.