lint/anybooru

---
license: openrail
---

# Anybooru

## Synthetic Anime Image Dataset

Synthetic anime image dataset generated using Andite's Anything-v4.5 checkpoint with Danbooru2021 tags collected in https://gwern.net/danbooru2021.
See https://github.com/1lint/anybooru for details and code to generate your own variant of the dataset. I have also uploaded the extracted Danbooru2021 tags to https://huggingface.co/datasets/lint/danbooru_tags, this dataset was generated with a small subset of the tags in `2021_0_pruned.parquet`.

Each string of tags was used to generate 4 different images with different seeds. This serves a similar purpose as random resize crop and image flip transformations to train the model to focus on general concepts encoded in the tags, rather than memorizing specific images.

## Quick Start

```
from datasets import load_dataset

dataset = load_dataset('lint/anybooru')

sample = dataset['train'][0]

image = sample['image']
tags = image.info['tags']

print(tags)
```

## Samples

Each row of samples share the same generation prompt (string of tags).


## Citations

```
@misc{danbooru2021, author = {Anonymous and Danbooru community and Gwern Branwen}, title = {Danbooru2021: A Large-Scale Crowdsourced and Tagged Anime Illustration Dataset}, howpublished = {\url{https://gwern.net/danbooru2021}}, url = {https://gwern.net/danbooru2021}, type = {dataset}, year = {2022}, month = {January}, timestamp = {2022-01-21}, note = {Accessed: 03/01/2023} }
```

原始信息汇总

Anybooru 合成动漫图像数据集概述

数据集描述

• 数据集名称：Anybooru
• 类型：合成动漫图像数据集
• 生成方法：使用Andite的Anything-v4.5模型进行生成，结合Danbooru2021标签。
• 标签来源：Danbooru2021标签，详情可访问https://gwern.net/danbooru2021。
• 自定义生成：提供代码和方法以生成数据集的变体，详情见https://github.com/1lint/anybooru。
• 标签数据集：已上传提取的Danbooru2021标签至https://huggingface.co/datasets/lint/danbooru_tags。

数据集特性

• 图像生成：每个标签字符串生成4张不同种子下的图像。
• 训练目的：通过不同图像训练模型关注标签中的通用概念，而非特定图像。

使用方法

• 加载数据集：使用datasets库加载数据集。
• 示例代码： python from datasets import load_dataset dataset = load_dataset(lint/anybooru) sample = dataset[train][0] image = sample[image] tags = image.info[tags] print(tags)

样本展示

• 样本结构：每行样本共享相同的生成提示（标签字符串）。
• 样本图像：展示于./anybooru_grid.png。

引用信息

• 引用格式：

  @misc{danbooru2021, author = {Anonymous and Danbooru community and Gwern Branwen}, title = {Danbooru2021: A Large-Scale Crowdsourced and Tagged Anime Illustration Dataset}, howpublished = {url{https://gwern.net/danbooru2021}}, url = {https://gwern.net/danbooru2021}, type = {dataset}, year = {2022}, month = {January}, timestamp = {2022-01-21}, note = {Accessed: 03/01/2023} }