_{^{TwinFlow: Realizing One-step Generation on Large Models with Self-adversarial Flows}}

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

Join the WeChat Group, feel free to reach out anytime if you have any questions!👇

👇 WeChat Group QR Code/微信群二维码 👇

• Technical Discussion Group 1 is full, please join Technical Discussion Group 2

🧭 Table of Contents

• 🔥 Codebase Usage 🔥
• Inference Demo
• Tutorials on MNIST

📰 News

• TwinFlow is accepted to ICLR 2026, see you in 🇧🇷!
• Thanks to @mengqin for adapting more compatible TwinFlow-models workflows in ComfyUI! 👏🏻
• Thanks to @smthemex for adapting TwinFlow-models workflows in ComfyUI! 👏🏻
• We release experimental version of faster Z-Image-Turbo!
• We release training code and better TwinFlow implementation on SD3.5 and OpenUni under src directory 👏🏻.
• 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!

⚙️ Key Features

💪 Open-source Plans

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

TwinFlow

TwinFlow-Z-Image-Turbo-exp Visualization

👀 Original Z-Image-Turbo 2-NFE

2-NFE visualization of Z-Image-Turbo

TwinFlow-Qwen-Image Visualization

Comparison with Qwen-Image and Qwen-Image-Lightning

---

Overview

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.

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.

Inference Demo

For ComfyUI users, please see https://github.com/smthemex/ComfyUI_TwinFlow.

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],
}

📖 Citation

@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}
}

@misc{sun2025anystep,
  author = {Sun, Peng and Lin, Tao},
  note   = {GitHub repository},
  title  = {Any-step Generation via N-th Order Recursive Consistent Velocity Field Estimation},
  url    = {https://github.com/LINs-lab/RCGM},
  year   = {2025}
}

@article{sun2025unified,
  title = {Unified continuous generative models},
  author = {Sun, Peng and Jiang, Yi and Lin, Tao},
  journal = {arXiv preprint arXiv:2505.07447},
  year = {2025},
  url = {https://arxiv.org/abs/2505.07447},
  archiveprefix = {arXiv},
  eprint = {2505.07447},
  primaryclass = {cs.LG}
}

🤗 Acknowledgement

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

Note: The LINs Lab has openings for PhD students for the Fall 2026/2027 intake. Interested candidates are encouraged to reach out.