Three-Body Scattering for Generative Modeling
The paper proposes a one-step generator trained by “three-body” interactions instead of a critic, diffusion path, or autoregressive split.
TBSM treats each generated projectile as pulled toward one real sample and pushed away from one independently generated sample. The authors say this gives direct regression supervision tied to the 2-Wasserstein gradient-flow velocity of an energy-distance objective. In ImageNet-256 experiments, they report FID 2.23 with PixelDiT-XL in pixel space and 1.63 with DiT-XL in latent space at NFE=1. HF Daily Papers' note
TBSM treats each generated projectile as pulled toward one real sample and pushed away from one independently generated sample. The authors say this gives direct regression supervision tied to the 2-Wasserstein gradient-flow velocity of an energy-distance objective. In ImageNet-256 experiments, they report FID 2.23 with PixelDiT-XL in pixel space and 1.63 with DiT-XL in latent space at NFE=1. HF Daily Papers' note
score 5