TBSM: Three-Body Scattering for Generative Modeling
_{Three-Body Interaction. One-Step Generation!}

Peng Sun^1,2 · Zhenglin Cheng^1,2 · Deyuan Liu¹ · Jun Xie^1,2 · Xinyi Shang³ · Tao Lin¹

¹Westlake University    ²Zhejiang University    ³University College London

Paper · Why TBSM? · Quick start · TODO · Citation

tbsm_demo.mp4

_{TBSM in two acts: from three-body scattering to one-step generation.}

_{One-step generation across datasets, resolutions, and output spaces.}

_{One-step text-to-image generation at 1024 × 1024 resolution.}

Why TBSM?

Structural bottleneck	Practical limitation
GANs learned adversarial critic	training is sensitive to the adversarial min–max objective
AR / Diffusion ordered factorization or path-indexed objective	sequential decoding or iterative sampling slows inference
Drifting Models minibatch-wide pairwise field	all-pairs cost grows with batch size same-condition batches poorly fit one-image-per-caption T2I data
Diffusion distillation often requires extra distillation losses, auxiliary training models, schedules, or stabilization tricks	some pipelines require additional teacher or online score networks during training, increasing systems complexity and potentially peak GPU memory

At its energy-distance endpoint, TBSM turns a proper distributional energy into direct, sample-level supervision for one-step generation—without an adversarial critic, teacher queries, or a batch-wide all-pairs field.

Two parameters, one design space.
_{The (ρ, λ) map links four TBSM configurations to Drift-like dynamics, diffusion-related supervision, and GAN-like objectives.}

Quick start

Performance

Resolution	Space	Backbone	NFE	FID ↓	IS ↑
256²	latent	DiT-XL/2	1	1.63	243.0
256²	pixel	PixelDiT-XL/16	1	2.23	230.9
256²	pixel	JiT-B/16	1	2.69	247.4
512²	latent	DiT-XL/4	1	1.92	252.1

Before and after · ImageNet-512

Before TBSM · NFE 1 · FID 398.20	After TBSM · NFE 1 · FID 1.92

Run

The minimal MNIST implementation provides a lightweight, low-cost path to understanding the core TBSM training algorithm through a compact, runnable example.

For full ImageNet-1K experiments, see the ImageNet training and evaluation code, including configurations for pixel- and latent-space generators.

TODO

• Release a minimal TBSM implementation for quick experimentation
• Release the ImageNet-1K training codebase
• Release ImageNet-1K checkpoints (by July 28, 2026)
• Release the Qwen-Image-20B checkpoint
• Release the Qwen-Image-20B training code
• Update the paper

Citation

@article{sun2026three,
  title   = {Three-Body Scattering for Generative Modeling},
  author  = {Sun, Peng and Cheng, Zhenglin and Liu, Deyuan and Xie, Jun and Shang, Xinyi and Lin, Tao},
  journal = {arXiv preprint arXiv:2607.18198},
  year    = {2026}
}

License

This is a mixed-license repository. Unless otherwise noted, original TBSM code is licensed under the Apache License 2.0. Vendored or adapted third-party files remain subject to their upstream licenses. In particular, imgnet/networks/pixdit_core/ is governed by the NVIDIA License (NSCLv1), which permits use only for non-commercial research or evaluation. See Third-Party Notices for file-level scope, attribution, and applicable license texts. External model weights and datasets remain subject to their upstream terms.