yangkaiSIGS/d3po_datasets

Datasets for the Direct Preference for Denoising Diffusion Policy Optimization (D3PO)

Description: This repository contains the dataset for the D3PO method in this paper Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model. The d3po_dataset file pertains to the image distortion experiment of the anything-v5 model. The text2img_dataset comprises the images generated from the pretrained, preferred image fine-tuned, reward weighted fine-tuned and D3PO fine-tuned models in the prompt-image alignment experiment.

Source Code: The code used to generate this data can be found here.

Directory

• d3po_dataset

  • epoch1
    • all_img
      • *.png
    • deformed_img
      • *.png
    • json
      • data.json (required for training)
    • prompt.json
    • sample.pkl(required for training)
  • epoch2`
  • ...
  • epoch5

• text2img_dataset:

  • img
  • data_*.json
  • plot.ipynb
  • prompt.txt

Citation

@article{yang2023using,
  title={Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model},
  author={Yang, Kai and Tao, Jian and Lyu, Jiafei and Ge, Chunjiang and Chen, Jiaxin and Li, Qimai and Shen, Weihan and Zhu, Xiaolong and Li, Xiu},
  journal={arXiv preprint arXiv:2311.13231},
  year={2023}
}