Three-Body Scattering for Generative Modeling
The paper proposes a one-step generator trained by “three-body” sample interactions instead of a critic, diffusion path, or autoregressive split.
Each generated projectile is pulled toward one real sample and pushed away from one independently generated sample. The authors say the expected interaction matches a Wasserstein gradient-flow velocity for an energy-distance objective. In image-feature space, the method trains one-step ImageNet-256 generators, reporting FID 2.23 with PixelDiT-XL and 1.63 with latent DiT-XL at NFE 1. ArXiv · AI/CL/LG's note
Each generated projectile is pulled toward one real sample and pushed away from one independently generated sample. The authors say the expected interaction matches a Wasserstein gradient-flow velocity for an energy-distance objective. In image-feature space, the method trains one-step ImageNet-256 generators, reporting FID 2.23 with PixelDiT-XL and 1.63 with latent DiT-XL at NFE 1. ArXiv · AI/CL/LG's note
score 5