🔥 News

• 2026-01-26: Our paper has been accepted by ICLR 2026. 🎉🎉🎉
• 2025-03-27: Add HuggingFace integration to ARPG.
• 2025-03-25: Add the sampling arccos schedule.
• 2025-03-14: The paper and code are released!

📖 Introduction

We introduce a novel autoregressive image generation framework named ARPG. This framework is capable of conducting BERT-style masked modeling by employing a GPT-style causal architecture. Consequently, it is able to generate images in parallel following a random token order and also provides support for the KV cache.

• 💪 ARPG achieves an FID of 1.94
• 🚀 ARPG delivers throughput 26 times faster than LlamaGen—nearly matching VAR
• ♻️ ARPG reducing memory consumption by over 75% compared to VAR.
• 🔍 ARPG supports zero-shot inference (e.g., inpainting and outpainting).
• 🛠️ ARPG can be easily extended to controllable generation.

🤗 Model Zoo

We provide the model weights pre-trained on ImageNet-1K 256*256.

Model	Param	Schedule	CFG	Step	FID	IS	Weight
ARPG-L	320 M	cosine	4.5	64	2.44	292	arpg_300m.pt
ARPG-XL	719 M	cosine	6.0	64	2.10	331	arpg_700m.pt
ARPG-XXL	1.3 B	cosine	7.5	64	1.94	340	arpg_1b.pt

🎮 Quick Start

You can easily play ARPG using the HuggingFace DiffusionPipeline.

from diffusers import DiffusionPipeline

pipeline = DiffusionPipeline.from_pretrained("hp-l33/ARPG", custom_pipeline="hp-l33/ARPG")

class_labels = [207, 360, 388, 113, 355, 980, 323, 979]

generated_image = pipeline(
    model_type="ARPG-XL",       # choose from 'ARPG-L', 'ARPG-XL', or 'ARPG-XXL'
    seed=0,                     # set a seed for reproducibility
    num_steps=64,               # number of autoregressive steps
    class_labels=class_labels,  # provide valid ImageNet class labels
    cfg_scale=4,                # classifier-free guidance scale
    output_dir="./images",      # directory to save generated images
    cfg_schedule="constant",    # choose between 'constant' (suggested) and 'linear'
    sample_schedule="arccos",   # choose between 'arccos' (suggested) and 'cosine'
)

generated_image.show()

If you want to train or reproduce the results of ARPG, please refer to Getting Started.

🔗 Bibtex

If this work is helpful for your research, please give it a star or cite it:

@inproceedings{li2026autoregressive,
    title={Autoregressive Image Generation with Randomized Parallel Decoding},
    author={Haopeng Li and Jinyue Yang and Guoqi Li and Huan Wang},
    booktitle={The Fourteenth International Conference on Learning Representations},
    year={2026}
}

🤝 Acknowledgement

Thanks to LlamaGen for its open-source codebase. Appreciate RandAR and RAR for inspiring this work, and also thank ControlAR.