What it is
Trains one-step generators by regressing each sample toward a real-attract, fake-repel scattering target from energy distance, avoiding critics and diffusion teachers. Instead of minibatch all-pairs fields or teacher-guided paths, it uses constant-size triplets: each generated sample moves toward one real sample and away from one fake. An online tracker estimates the conditional mean scattering field to reduce noise.
Key results
- ImageNet-256, NFE=1: FID 1.63 with latent-space DiT-XL
- ImageNet-256, NFE=1: FID 2.23 with pixel-space PixelDiT-XL
- ImageNet-512: DiT-XL/4 reaches FID 1.92 at NFE=1
Shows that energy-distance supervision can scale to high-dimensional one-step generation using only constant-size sample interactions.