Zhenglin Cheng^*   ·   Peng Sun^*   ·   Jianguo Li   ·   Tao Lin

• Inference Demo
• Tutorials on MNIST

• We release tutorials on MNIST to provide core implementation of TwinFlow!
• We release TwinFlow-Qwen-Image-v1.0! And we are also working on Z-Image-Turbo to make it faster!

1. Simple and Memory-Efficient Framework
   • No JVPs: Unlike sCM and MeanFlow, we require no JVP operations, making the method highly memory-friendly.
   • No GANs: Unlike DMD2, We eliminate the need for adversarial loss, removing the complexity of training discriminators.
   • No Auxiliary Networks: Unlike distribution matching methods like VSD/DMD, We require no fixed teacher models for distillation and no additional fake score networks for distribution matching.

This feature is demonstrated by our successful full-parameter few-step training of the Qwen-Image-20B.

2. Flexible Initialization: "Start with Any Model"
   • Further Distillation: Thanks to the one-model design, our method can learn the score function starting from any stage, enabling further distillation on already distilled models.

To verify this flexibility, try our method directly on the MNIST tutorials (random initialization) and stay tuned for our faster Z-Image-Turbo (distilled model initialization).

• Release inference and sampler code for TwinFlow-Qwen-Image-v1.0.
• Release training tutorials on MNIST for understanding.
• Release training code on SD3.5.
• Release faster Z-Image-Turbo.
• Release large-scale training code.

2-NFE visualization of TwinFlow-Qwen-Image

Case 1: 万里长城秋景，蜿蜒盘踞于层峦叠嶂的山脉之上，砖石城墙与烽火台在暖阳下呈现古朴的土黄色，山间枫叶如火般绚烂，游客点缀其间，远山薄雾缭绕，天空湛蓝飘着几朵白云，高角度全景构图，细节丰富，光影柔和。

Case2: 超高清壁纸, 梦幻光影, 少女在元宵灯会中回眸一笑, 提着一盏兔子花灯, 周围挂满明亮的灯笼, 暖色调灯光映照在脸上, 华丽的唐装, 繁复的头饰, 热闹的背景虚化, 焦外光斑美丽, 中景镜头。
Same prompt but different noise (left to right). Top to bottom shown are: Qwen-Image (50×2 NFE), TwinFlow-Qwen-Image (1-NFE), and Qwen-Image-Lightning-v2.0 (1-NFE).
TwinFlow-Qwen-Image generates high-quality images at 1-NFE while preserving strong diversity.

We introduce TwinFlow, a framework that realizes high-quality 1-step and few-step generation without the pipeline bloat.

Instead of relying on external discriminators or frozen teachers, TwinFlow creates an internal "twin trajectory". By extending the time interval to $t\in[−1,1]$, we utilize the negative time branch to map noise to "fake" data, creating a self-adversarial signal directly within the model.

Then, the model can rectify itself by minimizing the difference of the velocity fields between real trajectory and fake trajectory, i.e. the $\Delta_\mathrm{v}$. The rectification performs distribution matching as velocity matching, which gradually transforms the model into a 1-step/few-step generator.

TwinFlow method overview

Key Advantages:

• One-model Simplicity. We eliminate the need for any auxiliary networks. The model learns to rectify its own flow field, acting as the generator, fake/real score. No extra GPU memory is wasted on frozen teachers or discriminators during training.
• Scalability on Large Models. TwinFlow is easy to scale on 20B full-parameter training due to the one-model simplicity. In contrast, methods like VSD, SiD, and DMD/DMD2 require maintaining three separate models for distillation, which not only significantly increases memory consumption—often leading OOM, but also introduces substantial complexity when scaling to large-scale training regimes.

In the tutorials/mnist directory, we provide training code for MNIST that closely follows the implementation described in the paper, intended for tutorial purposes. This tutorial includes the core implementations of $L_\mathrm{base}$ and $\mathcal{L}_\mathrm{TwinFlow}$ .

To run TwinFlow training:

cd tutorials/mnist
python main.py --using_twinflow --save_dir ./outputs/twinflow

To run training without $\mathcal{L}_\mathrm{TwinFlow}$:

cd tutorials/mnist
python main.py --save_dir ./outputs/rcgm

TwinFlow training on MNIST	RCGM (without TwinFlow) training on MNIST

Install the latest diffusers:

pip install git+https://github.com/huggingface/diffusers

Run inference demo inference.py:

python inference.py

We recommend to sample for 2~4 NFEs:

# 4 NFE config
sampler_config = {
    "sampling_steps": 4,
    "stochast_ratio": 1.0,
    "extrapol_ratio": 0.0,
    "sampling_order": 1,
    "time_dist_ctrl": [1.0, 1.0, 1.0],
    "rfba_gap_steps": [0.001, 0.5],
}

# 2 NFE config
sampler_config = {
    "sampling_steps": 2,
    "stochast_ratio": 1.0,
    "extrapol_ratio": 0.0,
    "sampling_order": 1,
    "time_dist_ctrl": [1.0, 1.0, 1.0],
    "rfba_gap_steps": [0.001, 0.6],
}

@article{cheng2025twinflow,
  title={TwinFlow: Realizing One-step Generation on Large Models with Self-adversarial Flows},
  author={Cheng, Zhenglin and Sun, Peng and Li, Jianguo and Lin, Tao},
  journal={arXiv preprint arXiv:2512.05150},
  year={2025}
}

TwinFlow is built upon RCGM and UCGM, with much support from InclusionAI.