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](https://arxiv.org/abs/2311.13231). The *d3po_dataset* file pertains to the image distortion experiment of the [`anything-v5`](https://huggingface.co/stablediffusionapi/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](https://github.com/yk7333/D3PO/).

**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}
}
```

原始信息汇总

D3PO 数据集

描述

该数据集用于论文《Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model》中的 D3PO 方法。数据集包括以下两个部分：

• d3po_dataset: 与 anything-v5 模型的图像失真实验相关。
• text2img_dataset: 包含从预训练、偏好图像微调、奖励加权微调和 D3PO 微调模型生成的图像，用于提示-图像对齐实验。

目录结构

• d3po_dataset:

  • epoch1
    • all_img
      • *.png
    • deformed_img
      • *.png
    • json
      • data.json (训练所需)
    • prompt.json
    • sample.pkl (训练所需)
  • epoch2
  • ...
  • epoch5

• text2img_dataset:

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

引用

@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} }