sp12138/TBSM

TL;DR: TBSM is a new training principle for one-step generative models from Westlake University, framed as "three-body scattering": a proper energy-distance objective reduced to direct, sample-level supervision — no adversarial critic, no teacher network, no diffusion sampling chain, no batch-wide all-pairs interactions. Self-reported one-step results are startlingly good: FID 1.63 on ImageNet-256 (latent), 2.23 in pixel space, and 1.92 at 512², all at NFE=1, with ImageNet training code released. The paper is days old and the flagship 1024² text-to-image checkpoint (Qwen-Image-20B) is unreleased, so treat the numbers as unverified.

What it claims to be

The pitch is that every existing route to fast generation has a structural bottleneck, and the README lays them out explicitly: GANs need an adversarial critic and a fragile min–max game; diffusion/AR models pay for their path-indexed or ordered objectives with iterative sampling; distillation needs teacher queries, auxiliary losses, and extra GPU memory; and Drifting Models (the recent minibatch-field approach) compute an all-pairs interaction across the batch, which scales badly and fits poorly to one-caption-per-image T2I data.

TBSM's answer: start from an energy distance — a proper scoring rule between distributions — and at its endpoint the objective collapses into per-sample supervision computed from small groups of interacting samples (the "three bodies"), rather than a learned critic or a batch-wide field. The README also describes a two-parameter (ρ, λ) design space in which TBSM configurations interpolate between Drift-like dynamics, diffusion-related supervision, and GAN-like objectives — i.e., the known methods appear as points on the same map. That unifying framing, if the paper backs it up, is the genuinely interesting part; the README itself doesn't spell out the loss, so the mechanism details below the "three-body" slogan live in the arXiv paper.

Distillation teacher net extra model in memory

Drifting all-pairs, O(batch²)

TBSM three-body scattering

Energy-distance endpoint → direct sample-level supervision for one-step generation no critic · no teacher · no batch-wide field · fits one-caption-per-image T2I data

The numbers

All results are one-step (NFE=1), on ImageNet, self-reported:

If real, 1.63 FID in a single forward pass beats the original DiT-XL/2 running 250 steps, and 2.23 in raw pixel space (no VAE) at one step would be remarkable — pixel-space one-step generation has historically been much worse than latent. The "before/after" figure (FID 398 → 1.92 at 512²) shows the training does the heavy lifting, though the README doesn't clarify whether "before" is a random init or a pretrained diffusion model evaluated at one step — worth checking in the paper, since it determines whether this is training-from-scratch or a fine-tuning recipe.

There's also a 1024×1024 text-to-image demo built on Qwen-Image-20B — but that checkpoint and its training code are the two unchecked TODO items.

What's actually released

• Minimal MNIST implementation — a compact, runnable version of the core training algorithm. This is the right place to understand the loss.
• Full ImageNet-1K training and eval code, with configs for both pixel- and latent-space generators, plus training assets on Hugging Face.
• Not released: the Qwen-Image-20B T2I checkpoint and training code; the paper itself is flagged for an update.

License caveat: mostly Apache-2.0, but the vendored PixelDiT backbone is under NVIDIA's NSCLv1 (non-commercial research only), so the pixel-space configs are not commercially usable as-is.

Caveats

Preprint is ~2 weeks old, 71 stars, no independent reproductions yet, and all numbers are self-reported. The README leans on physics metaphor without stating the actual objective, so verify the "no tricks needed" claim against the paper — one-step training methods have a long history of hidden stabilization requirements. The most impressive artifact (1024² T2I) is exactly the one not yet released.

Try it

The README gives no install commands; start from the MNIST example:

git clone https://github.com/sp12138/TBSM
cd TBSM
# Minimal TBSM: see the MNIST implementation linked from Quick start
# Full-scale: ImageNet-1K training/eval configs (pixel & latent) in the repo,
# with training assets on Hugging Face (linked in the TODO section)

Paper: arXiv:2607.18198. Worth a close read of the loss derivation and the (ρ, λ) unification — if the framing holds, this is a new family of one-step objectives, not another distillation variant.