OSPO

This repository contains the official implementation for the paper
"OSPO: Object-centric Self-improving Preference Optimization for Text-to-Image Generation."

0. About

OSPO is a self-improving prefernece optimization framework for compositional text-to-image generation, allowing an MLLM to improve its fine-grained image generation capability without needing any data or external model.

TODO

• Release the model checkpoint.
• Release the training dataset.
• Release the inference code for text-to-image generation.
• Rlease the OSPO framework code.
• Release the evaluation code.
• Release the Unitok version code.
• Release example data for each step.

1. Setup

1. Create Conda Environment

conda create -n ospo python=3.10 -y
conda activate ospo

2. Clone this repository

git clone https://github.com/KU-AGI/OSPO.git
cd OSPO

3. Install Dependencies

conda install pytorch==2.2.1 torchvision==0.17.1 torchaudio==2.2.1 pytorch-cuda=12.1 -c pytorch -c nvidia

# Install Dependencies
pip install -r requirements.txt

4. Download Janus-Pro-7B Model (baseline) from 🤗 HuggingFace

• Create ./checkpoints directory and place the Model in the ./checkpoints directory.
• So the Model folder path should be ./checkpoints/Janus-Pro-7B

2. Inference

We provide our different model checkpoints on Hugging Face.

Model	Download
OSPO + Janus-Pro-7B	🤗 HuggingFace
OSPO + Janus-1.3B	🤗 HuggingFace
OSPO + Unitok-MLLM-7B	🤗 HuggingFace

1. Download the model weights from the table.
2. Place the checkpoint in the ./checkpoints directory.

3. Run the code.

# OSPO + Janus-Pro-7B
python inference_janus.py --input ${INPUT_TEXT}

# OSPO + Unitok-MLLM-7B
python inference_unitok.py --input ${INPUT_TEXT}

Results will be saved in ./results

3. Reproduce

First of all, you can reproduce only the training step using our provided training dataset. 🤗 HuggingFace

python ospo/step4.py --cfg_path configs/step4.yaml

If you want to reproduce the full OSPO framework, please refer to this page.

4. Evaluate

Please refer to this page.

📍 Acknowledgement

We thank the authors of Janus-Pro, UniTok, SimPO and mDPO, for making their code available.

📍 Citation

@article{oh2025object,
  title={Object-centric Self-improving Preference Optimization for Text-to-Image Generation},
  author={Oh, Yoonjin and Kim, Yongjin and Kim, Hyomin and Chi, Donghwan and Kim, Sungwoong},
  journal={arXiv preprint arXiv:2506.02015},
  year={2025}
}