|
--- |
|
license: mit |
|
task_categories: |
|
- image-classification |
|
- image-to-image |
|
- text-to-image |
|
language: |
|
- en |
|
- ja |
|
pretty_name: danbooru2023 |
|
size_categories: |
|
- 1M<n<10M |
|
viewer: false |
|
--- |
|
|
|
<img src="https://huggingface.co/datasets/nyanko7/danbooru2023/resolve/main/cover.webp" alt="cover" width="750"/> |
|
|
|
# Danbooru2023: A Large-Scale Crowdsourced and Tagged Anime Illustration Dataset |
|
|
|
<!-- Provide a quick summary of the dataset. --> |
|
|
|
Danbooru2023 is a large-scale anime image dataset with over 5 million images contributed and annotated in detail by an enthusiast community. Image tags cover aspects like characters, scenes, copyrights, artists, etc with an average of 30 tags per image. |
|
|
|
Danbooru is a veteran anime image board with high-quality images and extensive tag metadata. The dataset can be used to train image classification, multi-label tagging, character detection, generative models, and other computer vision tasks. |
|
|
|
- **Shared by:** Nyanko Devs |
|
- **Language(s):** English, Japanese |
|
- **License:** MIT |
|
|
|
This dataset is built on the top of [danbooru2021](https://gwern.net/danbooru2021). We expands the dataset to include images up to ID #6,857,737, adding over 1.8 million additional images and total size is now approximately 8 terabytes (8,000 GB). |
|
|
|
## Use |
|
|
|
## Format |
|
|
|
The goal of the dataset is to be as easy as possible to use immediately, avoiding obscure file formats, while allowing simultaneous research & seeding of the torrent, with easy updates. |
|
|
|
Images are provided in the full original form (be that JPG, PNG, GIF or otherwise) for reference/archival purposes, and bucketed into 1000 subdirectories 0000–0999 (0-padded), which is the Danbooru ID modulo 1000 (ie. all images in 0999/ have an ID ending in ‘999’); IDs can be turned into paths by dividing & padding (eg. in Bash, BUCKET=$(printf "%04d" $(( ID % 1000 )) )) and then the file is at {original,512px}/$BUCKET/$ID.$EXT. |
|
|
|
The reason for the bucketing is that a single directory would cause pathological filesystem performance, and modulo ID is a simple hash which spreads images evenly without requiring additional future directories to be made or a filesystem IO to check where the file is. The ID is not zero-padded and files end in the relevant extension, hence the file layout looks like this: |
|
|
|
```bash |
|
$ tree / | less |
|
|
|
/ |
|
├── danbooru2023 -> /mnt/diffusionstorage/workspace/danbooru/ |
|
│ ├── metadata |
|
│ ├── readme.md |
|
│ ├── original |
|
│ │ ├── 0000 -> data-0000.tar |
|
│ │ ├── 0001 -> data-0001.tar |
|
│ │ │ ├── 10001.jpg |
|
│ │ │ ├── 210001.png |
|
│ │ │ ├── 3120001.webp |
|
│ │ │ ├── 6513001.jpg |
|
|
|
``` |
|
|
|
Currently represented file extensions are: avi/bmp/gif/html/jpeg/jpg/mp3/mp4/mpg/pdf/png/rar/swf/webm/wmv/zip. |
|
|
|
Raw original files are treacherous. Be careful if working with the original dataset. There are many odd files: truncated, non-sRGB colorspace, wrong file extensions (eg. some PNGs have .jpg extensions like original/0146/1525146.jpg or original/0558/1422558.jpg), etc. |
