Three-Body Scattering for Generative Modeling

The paper introduces Three-Body Scattering Modeling (TBSM), which converts a proper distributional energy into sample-level motion for training one-step generators. Each projectile is attracted to one real source and repelled from one independently generated source. The conditional expectation of this interaction is claimed to equal the Wasserstein-2 gradient-flow velocity of one-half the squared energy distance. Unlike minibatch-wide all-pairs fields, a batch of B frozen-target events produces O(B) losses, using one reference per condition. With scattering computed in frozen image features, TBSM reports ImageNet-256 FID scores of 2.23 for pixel-space PixelDiT-XL and 1.63 for latent-space DiT-XL at NFE=1.