ReasonGen-R1:
Cot for Autoregressive Image generation models through SFT and RL

📥 Model and Dataset Download | ⚡ Quick Start | 📜 Acknowledgement | 📖 Citation
📄 Paper Link

1. Introduction

Although chain-of-thought (CoT) reasoning and reinforcement learning (RL) have driven breakthroughs in NLP, their integration into generative vision models remains underexplored. We introduce ReasonGen-R1, a two-stage framework that first imbues an autoregressive image generator with explicit text-based "thinking" skills via supervised fine-tuning (SFT) on a newly generated reasoning dataset of written rationales, and then refines its outputs using Group Relative Policy Optimization (GRPO). To enable the model to reason through text before generating images, We automatically generate and release a corpus of model-crafted rationales paired with visual prompts, enabling controlled planning of object layouts, styles, and scene compositions. Our GRPO algorithm uses reward signals from a pretrained vision–language model to assess overall visual quality, optimizing the policy in each update. Evaluations on Geneval, DPG, and the T2I benchmark demonstrate that ReasonGen-R1 consistently outperforms strong baselines and prior state-of-the-art models. We will open-source our generated reasoning dataset and training code to accelerate further advances in text-based reasoning–driven image generation.

2. Model and Dataset Download

Huggingface

Model	Download
ReasonGen-R1	🤗 Hugging Face
ReasonGen-R1-SFT-Only	🤗 Hugging Face

Dataset	Download
ReasonGen-R1-Datasets	🤗 Hugging Face

3. Quick Start

Installation

You can install the necessary dependencies by running the following command:

cd ~
mkdir project
cd project
conda create -n image_rl python==3.12 -y
conda activate image_rl
pip3 install torch==2.6.0 torchvision --index-url https://download.pytorch.org/whl/cu124
pip3 install flash-attn==2.7.4.post1 --no-build-isolation
git clone https://github.com/Franklin-Zhang0/ReasonGen-R1.git
cd ReasonGen-R1
pip install -r requirements.txt
pip install -e .
pip install -e ./Janus

Evaluation Environment Installation (Optional)

If you want to run the evaluation code, you can install the evaluation environment by running the following commands:

# Geneval
cd ~
mkdir project
cd project
git clone https://github.com/djghosh13/geneval.git
cd geneval
conda create -n geneval python=3.9 -y
conda activate geneval
pip install torch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1
pip install mmcv-full==1.7.0 -f https://download.openmmlab.com/mmcv/dist/cu117/torch1.13/index.html
pip install mmengine==0.7.3

pip install pandas
pip install numpy==1.23.1

pip install open-clip-torch
pip install clip-benchmark

git clone https://github.com/open-mmlab/mmdetection.git
cd mmdetection; git checkout 2.x
pip install -v -e .

cd ../
bash ./evaluation/download_models.sh "./models"

# DPG
cd ~
cd project
git clone https://github.com/TencentQQGYLab/ELLA.git
cd ELLA
cp ~/project/ReasonGen-R1/benchmark/requirements-for-dpg_bench.txt .
conda create -n dpg_test python=3.9 -y
conda activate dpg_test
conda install conda-forge::fairseq -y
pip install -r requirements-for-dpg_bench.txt

Once the eval environment is setup, you can use the following commands to run the evaluation:

bash -i benchmark/geneval.sh
bash -i benchmark/dpg_eval.sh

Inference

To inference with the ReasonGen-R1 model, you can use the following command:

python ReasonGen-R1/Janus/cot_generate_inference.py

SFT Training

To train the SFT model from Janus-Pro-7B model on the ReasonGen-R1-SFT-200k dataset, you can use the following command:

bash ReasonGen-R1/examples/janus_sft.sh

RL Training

To train the RL model from the ReasonGen-R1-SFT model, you can use the following command:

bash ReasonGen-R1/Janus/janus_rl.sh

4. Acknowledgements

We would like to thank Verl, upon which our repo is built.

5. Citation

@misc{zhang2025reasongenr1cotautoregressiveimage,
      title={ReasonGen-R1: CoT for Autoregressive Image generation models through SFT and RL}, 
      author={Yu Zhang and Yunqi Li and Yifan Yang and Rui Wang and Yuqing Yang and Dai Qi and Jianmin Bao and Dongdong Chen and Chong Luo and Lili Qiu},
      year={2025},
      eprint={2505.24875},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2505.24875}, 
}

6. More Examples