NAMI: Efficient Image Generation via Bridged Progressive Rectified Flow Transformers

The proposed NAMI architecture reduces the inference time required to generate 1024 resolution images by 64%, while still maintaining a high level of image quality.

💡 Update

• [2026.03.06] Release the NAMI-2B inference code and weights.

• [2026.03.01] Paper was accepted by CVPR2026.

🧩 Environment Setup

1、pip install -r requirements.txt
2、Directly use our diffusers folder, or replace the corresponding files in the installed diffusers package in your Python environment with transformer_flux.py and pipeline_flux.py from the src directory.

📂 Preparation of Model Weights

We provide model weights for evaluation and deployment. Please download files from NAMI, mclip, mt5-xxl and place them in the weights directory.

⏳ Inference Pipeline

Here we provide the inference demo for our NAMI.

cd src
python infer.py

🌸 Acknowledgement

This code is mainly built upon Diffusers, Flux, and PyramidFlow repositories. Thanks so much for their solid work!

💖 Citation

If you find this repository useful, please consider citing our paper:

@article{ma2025nami,
  title={NAMI: Efficient Image Generation via Bridged Progressive Rectified Flow Transformers},
  author={Ma, Yuhang and Cheng, Bo and Liu, Shanyuan and Zhou, Hongyi and Wu, Liebucha and Leng, Dawei and Yin, Yuhui},
  journal={arXiv preprint arXiv:2503.09242},
  year={2025}
}