Researchers from Hugging Face have introduced Three-Body Scattering Modeling (TBSM), a novel generative framework that learns a transport field from generated samples toward real data. Unlike GANs, diffusion models, or distillation methods, TBSM uses a tracked scattering field to directly guide the generator, enabling stable one-step and few-step training without auxiliary losses.
TBSM turns the energy distance into a constant-size per-projectile interaction: each projectile is attracted toward one real source and repelled from one independently generated source. Conditioned on the projectile and its condition, its expectation equals the 2-Wasserstein gradient-flow velocity of ½ D_E²(P_θ, Q). A batch of B frozen-target events yields O(B) sample-level losses, each using one reference for its condition instead of the minibatch-wide all-pairs field used by methods such as Drifting Models.
On ImageNet-256, TBSM-trained one-step generators achieve competitive generation quality with NFE=1: FID 2.23 with pixel-space PixelDiT-XL and FID 1.63 with latent-space DiT-XL. Beyond standard image generation, TBSM demonstrates stable training of large-scale text-to-image models with up to 20B parameters, achieving one-step and few-step generation without additional auxiliary losses.
TBSM also provides a design map that reveals connections among GANs, diffusion models, and drifting-based approaches, offering a new perspective on generative modeling paradigms. The code is available on GitHub at https://github.com/sp12138/TBSM.