🔈 News

• [2025-09] All weights are avialable on Huggingface!
• [2025-09] Train and inference code are avialable
• [2025-09] GCPO is released on Arixv.

🔍 Introduction

We propose a novel reinforcement learning framework, Group Critical-token Policy Optimization (GCPO), to achieve efficient policy optimization of critical tokens. We consider critical tokens from three perspectives:

• Causal dependency
• Entropy-induced spatial structure
• RLVR-focused token diversity

We select 30% of critical tokens from all tokens and combine them with Dynamic Advantage Weights to achieve precise optimization✌

We validate the effectiveness of GCPO on multiple models (LlamaGen, Janus-Pro) and text-to-image benchmarks (Geneval, T2I-CompBench, DrawBench). Notably, GCPO achieves 0.90 on Geneval built on the original Janus-Pro-7B, and has strong genvalidate the effectiveness of GCPO on multiple models (LlamaGen, Janus-Pro) and text-to-image benchmarks (Geneval, T2I-CompBench, DrawBench). Notably, GCPO achieves 0.90 on Geneval built on the original Janus-Pro-7B, and has strong generalization on T2I-CompBench.

CLICK for Detailed Results

Visualization Results

Geneval Results

T2I-CompBench Results

🤗 Model List

🔧 Environment SetUp

1. Clone this repository and navigate to the folder:

git clone https://github.com/zghhui/GCPO.git
cd GCPO

2. Install the training package:

We provide training codes for LlamaGen and Janus-Pro, and recommend installing the environments for each.

For LlamaGen:

conda create -n gcpo_llamagen python=3.10
conda activate gcpo_llamagen
pip install -r llamaGen/requirements.txt

For Janus-Pro:

conda create -n gcpo_janus python=3.10
conda activate gcpo_janus
pip install -r janus/requirements.txt

3. Download Models

For LlamaGen:

huggingface-cli download zghhui/LlamaGen-T2I
huggingface-cli download google/flan-t5-xl
wget https://huggingface.co/peizesun/llamagen_t2i/resolve/main/vq_ds16_t2i.pt

For Janus-Pro:

huggingface-cli download deepseek-ai/Janus-Pro-1B
huggingface-cli download deepseek-ai/Janus-Pro-7B

For HPS Reward:

huggingface-cli xswu/HPSv2

For Geneval Reward:

• Please according to the instructions in Flow-GRPO and reward-server.

🚀 Training GCPO

LlamaGen

cd llamaGen/src
bash scripts/rl_gcpo_hps.sh

[!Note]

Remember to modify the t5_model path in gcpo/llamaGen/simpar/model/llama_model.py (line 1244)

Janus-Pro

cd janus/src
bash scripts/run_gcpo_hps.sh
bash scripts/run_gcpo_geneval.sh

[!Note]

Please run geneval server before running geneval reward. The reward function is located in utils/reward_geneval.py, and the IP of server can be modified here.

💫 Inference

LlamaGen

cd llamaGen
bash scripts/inference.sh

Janus-Pro

cd janus/src
bash scripts/inference.sh

📧 Contact

If you have any comments or questions, please open a new issue.

🤗 Acknowledgments

Our training code is based on T2I-R1, SimpleAR, and Flow-GRPO.

Thanks to all the contributors!

⭐ Citation

If you find GCPO useful for your research or projects, we would greatly appreciate it if you could cite the following paper:

@article{zhang2025group,
  title={Group Critical-token Policy Optimization for Autoregressive Image Generation},
  author={Zhang, Guohui and Yu, Hu and Ma, Xiaoxiao and Zhang, JingHao and Pan, Yaning and Yao, Mingde and Xiao, Jie and Huang, Linjiang and Zhao, Feng},
  journal={arXiv preprint arXiv:2509.22485},
  year={2025}
}