Resurrect Mask AutoRegressive Modeling for Efficient and Scalable Image Generation

📚 Introduction

• In this work, we propose an enhanced mask autoregressive model architecture, Masked Generative Image LLaMA (MaskGIL).
• Building upon this architecture, we introduce a series of class-driven and text-driven image generation models with parameter sizes ranging from 111M to 1.4B.
• Beyond its core image generation capabilities, we further extend MaskGIL to a broader range of applications, including: (1) accelerating AR-based generation and (2) developing a real-time speech-to-image generation system.

🔥 News

[2025.7.17] 🎉🎉🎉 We released the technical report on arXiv.

[2025.4.10] 🎉🎉🎉 Code and Class-driven checkpoints are released!

📽️ Demo Examples

Class-Driven Generation Results

Text-Driven Generation Results

Unified AR and MAR Generation Results

💻 Repository Structure

Here's an overview of the repository structure:

  ├ MaskGIL/
  |    ├── Metrics/                               <- evaluation tool
  |    |      ├── inception_metrics.py                  
  |    |      └── sample_and_eval.py
  |    |    
  |    ├── Network/                             
  |    |      ├── Taming/                         <- VQGAN architecture   
  |    |      ├── tokenizer/                      <- VQGAN architecture  
  |    |      ├── gpt.py                          <- Bi-Direction LLaMA architecture      
  |    |      └── transformer.py                  <- Bi-Transformer architecture  
  |    |
  |    ├── Trainer/                               <- Main class for training
  |    |      ├── trainer.py                      <- Abstract trainer     
  |    |      └── vit.py                          <- Trainer of MaskGIL
  |    ├── images/                                <- Image samples         
  |    |
  |    ├── requirements.yaml                      <- help to install env 
  |    ├── FID_sample.py                          <- sample 50K images for FID
  |    └── main.py                                <- Main

🚀 Quick Start

1. Inference

python FID_sample.py

2. Training

bash train.sh

🤗 Checkpoints

Method	params	tokens	FID (256x256)	weight
MaskGIL-B	111M	16x16	5.64	c2i_B_256.pt
MaskGIL-L	343M	16x16	4.01	c2i_L_256.pt
MaskGIL-XL	775M	16x16	3.90	c2i_X_256.pt
MaskGIL-XXL	1.4B	16x16	3.71	c2i_XXL_256.pt

📖 BibTeX

@article{xin2025resurrect,
  title={Resurrect Mask AutoRegressive Modeling for Efficient and Scalable Image Generation},
  author={Xin, Yi and Zhuo, Le and Qin, Qi and Luo, Siqi and Cao, Yuewen and Fu, Bin and He, Yangfan and Li, Hongsheng and Zhai, Guangtao and Liu, Xiaohong and others},
  journal={arXiv preprint arXiv:2507.13032},
  year={2025}
}