Monetico: An Efficient Reproduction of Meissonic for Text-to-Image Synthesis

🚀 Introduction

Similar to Meissonic, Monetico is a non-autoregressive masked image modeling text-to-image synthesis model capable of generating high-resolution images. It is designed to run efficiently on consumer-grade graphics cards.

Monetico is an efficient reproduction of Meissonic. Trained on 8 H100 GPUs for approximately one week, Monetico can generate high-quality 512x512 images that are comparable to those produced by Meissonic and SDXL.

Monetico was developed by Collov Labs. We extend our gratitude to @MeissonFlow and @viiika for their valuable advice on efficient training.

🛠️ Prerequisites

Step 1: Clone the repository

git clone https://github.com/collovlabs/Monetico/
cd Monetico

Step 2: Create virtual environment

conda create --name monetico python
conda activate monetico
pip install -r requirements.txt

Step 3: Install diffusers

git clone https://github.com/huggingface/diffusers.git
cd diffusers
pip install -e .

💡 Usage

Gradio Web UI

python app_Monetico.py

Command-line Interface

Text-to-Image Generation

python inference_fp16_Monetico.py --prompt "Your creative prompt here"

📚 Citation

If you find this work helpful, please consider citing:

@article{bai2024meissonic,
  title={Meissonic: Revitalizing Masked Generative Transformers for Efficient High-Resolution Text-to-Image Synthesis},
  author={Bai, Jinbin and Ye, Tian and Chow, Wei and Song, Enxin and Chen, Qing-Guo and Li, Xiangtai and Dong, Zhen and Zhu, Lei and Yan, Shuicheng},
  journal={arXiv preprint arXiv:2410.08261},
  year={2024}
}

Made with ❤️ by the Collov AI