TBSM Enables Single-Step Image Generation, Accelerating 20B Text-to-Image Models

机器之心 · wechat · 2026-08-10

A new framework called TBSM (Three-Body Scattering), proposed by a joint PhD student from Westlake University and Zhejiang University, efficiently converts existing diffusion or flow-matching models into single-step generators (NFE=1).

Core Mechanism

Three-Body Scattering: Instead of relying on a fixed online teacher model, it constructs a local Fake-to-Real transfer direction using "one real sample + two generated samples."
Lightweight Tracker: A lightweight network (<30M parameters) is trained online to learn the stable mean of these scattering directions, guiding an 800M parameter generator to regress to target positions. The Tracker is discarded after training.
No Adversarial Training: Uses standard regression updates, avoiding the highly sensitive minimax optimization of GANs and eliminating the need for noise-time supervision used in traditional distillation.

Experimental Results

Achieves FID scores of 1.63 (latent space) and 2.23 (pixel space) for single-step generation on ImageNet-256.
Reduces the FID of a multi-step DiT-XL/4 model from 398.20 (when forced to single-step) down to 1.92 after TBSM training.
Successfully applied to a 20B parameter text-to-image model (initialized from Qwen-Image), achieving high-quality one-step generation without CFG.

The project has open-sourced MNIST and ImageNet-1K training code, with 20B model weights to be released later.

Original post →

TBSM Enables Single-Step Image Generation, Accelerating 20B Text-to-Image Models

More from Multimodal