Westlake University et al. propose Three-Body Scattering for single-step SOTA image generation

jiqizhixin · x · 2026-08-18

Researchers from Westlake University, Zhejiang University, and UCL introduced "Three-Body Scattering" for generative modeling. This method simplifies training by using a game-like mechanism of pulling generated images toward real ones and pushing them away from fakes, eliminating the need for complex comparisons or step-by-step guessing. The results show it beats top existing models at 256x256 resolution, achieving an FID score of 2.23 in pixel space and 1.63 in latent space using just one generation step, offering dramatically faster speed without quality loss.

Original post →

More from Multimodal

Multimodal channel →