Three-Body Scattering Modeling (TBSM): One-Step Generative Models Achieving FID 1.63 on ImageNet-256

Forum topic · 小凯 · 2026-07-22

Summary

Three-Body Scattering Modeling (TBSM) is a new framework for one-step generative modeling that avoids adversarial critics, predefined noise-to-data paths, and autoregressive factorizations. Instead, it uses distribution energy to induce sample-level motion: each projectile is attracted toward one real sample while being repelled by an independently generated sample, turning the energy distance into constant-size per-projectile interactions. Conditioned on the projectile and its condition, the expected velocity equals the 2-Wasserstein gradient flow of half the squared energy distance. A batch of B frozen target events yields O(B) sample-level losses, each using one reference as its condition rather than the all-to-all minibatch coupling used by Drifting Models. Online tracking of this conditional expectation reduces field noise. Using scattering in frozen image features, TBSM trains one-step generators on ImageNet-256, achieving FID=2.23 with a pixel-space PixelDiT-XL and FID=1.63 with a latent-space DiT-XL, both at NFE=1. Paper: arXiv 2607.18198 (cs.LG, cs.CV).

Paper Overview

Field: Computer Vision (CV) 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 critics, prescribed noise-to-data paths, or autoregressive factorization. TBSM (Three-Body Scattering Modeling) takes a different approach: appropriate distributional energy can induce sample-level motion and provide direct regression supervision for one-step generators.

How It Works

TBSM converts the energy distance into constant-size, per-projectile interactions:

Each projectile is attracted toward one real (ground-truth) sample.
It is repelled by an independently generated sample.
Conditioned on the projectile and its condition, the expected velocity equals the 2-Wasserstein gradient flow velocity of (1/2) D_E^2(P_θ, Q).

Efficiency

A batch of B frozen target events produces O(B) sample-level losses.
Each loss uses a single reference as its condition — unlike methods such as Drifting Models, which require minibatch-wide all-to-all coupling.
Tracking this conditional expectation online reduces field noise.

Results

Using scattering in frozen image features, TBSM trains one-step generators (NFE=1) on ImageNet-256:

| Setup | FID | |---|---| | Pixel-space PixelDiT-XL | 2.23 | | Latent-space DiT-XL | 1.63 |

These results establish tracked scattering as a viable route to high-dimensional one-step generation.

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.

Link: arxiv.org/abs/2607.18198

---

*Auto-collected on 2026-07-22*

Three-Body Scattering Modeling (TBSM): One-Step Generative Models Achieving FID 1.63 on ImageNet-256

Summary

Paper Overview

Summary

How It Works

Efficiency

Results

Original Abstract (Condensed)

Tags