Overview
Research area: Computer Vision / Machine Learning
Authors: Peng Sun, Zhenglin Cheng, Deyuan Liu, Jun Xie, Xinyi Shang, et al. (6 authors)
Published: 2026-07-20
arXiv: 2607.18198
Categories: cs.LG, cs.CV
Summary
Modern generative models typically rely on adversarial judges, predefined noise-to-data paths, or autoregressive factorization. TBSM instead demonstrates that a suitable distributional energy can induce sample-level motion and provide direct regression supervision for one-step generators.
Key Ideas
- Three-Body Scattering Modeling (TBSM) converts energy distances into constant-size per-projectile interactions: each projectile is attracted toward one real source and repelled by an independently generated source.
- Conditioned on the projectile and its target, the expected velocity equals a 2-Wasserstein gradient flow with (1/2) D_E²(P_θ, Q), where D_E is the energy distance.
- A batch of B frozen target events produces O(B) sample-level losses; each loss uses a single reference as its condition, avoiding the full all-to-all minibatch interactions used by methods like Drifting Models.
- Tracking this conditional expectation online reduces field noise.
- FID 2.23 with pixel-space PixelDiT-XL
- FID 1.63 with latent-space DiT-XL
Results
Using scattering in frozen image features, TBSM trains one-step generators on ImageNet-256 at NFE=1:
Original Abstract (Condensed)
> We propose Three-Body Scattering Modeling (TBSM) for one-step generation, achieving FID=2.23 with PixelDiT-XL and FID=1.63 with DiT-XL on ImageNet-256 at NFE=1, establishing tracked scattering as a route to high-dimensional one-step generation.
---
*Auto-collected on 2026-07-22*