README.md

---
license: mit
task_categories:
- image-classification
- image-to-image
- text-to-image
language:
- en
- ja
tags:
- anime
- gelbooru
- booru
- art
- webp
- deduplicated
pretty_name: Gelbooru2026
size_categories:
- 1M<n<10M
viewer: false
---

# Gelbooru2026: A Deduplicated Anime Illustration Dataset

## Dataset Description

Gelbooru2026 is a large-scale, community-tagged anime illustration dataset for
research and development in image generation, image classification, multimodal
learning, and related tasks.

This release contains **3,784,478 static images** distributed across 1,000 tar
shards. It was frozen in August 2026 and was processed using the same image
release format as
[Danbooru2026](https://huggingface.co/datasets/nyanko-devs/danbooru2026).

**Key specifications:**

* **Shared by:** Nyanko Devs
* **Source:** Gelbooru community
* **Language(s):** English, Japanese, mixed
* **Dataset license:** MIT; original images retain their respective copyrights
* **Image format:** WebP, quality 95, method 6
* **Image count:** 3,784,478
* **Shards:** 1,000 deterministic tar files

The dataset may contain adult or otherwise sensitive material. Users are
responsible for applying filtering appropriate to their use case.

## Deduplication Against Danbooru2026

This release is content-deduplicated against the released
[Danbooru2026](https://huggingface.co/datasets/nyanko-devs/danbooru2026)
dataset. Deduplication used two exact-content stages:

1. **9,950,543** Gelbooru candidates were excluded by matching their original
   file MD5 against a released Danbooru2026 image.
2. **25,474** additional candidates were excluded by matching the processed
   WebP SHA-256 against a released Danbooru2026 WebP.

In total, **9,976,017** candidates were removed by cross-dataset
deduplication. The final validation found **zero known processed SHA-256
intersections** with the released Danbooru2026 images.

This is exact byte/content deduplication, not perceptual or semantic
deduplication. Near-duplicates, crops, edits, recompressions, and visually
similar images can therefore remain.

## Image Processing

Gelbooru2026 contains static images only. Animations, videos, corrupted files,
unsupported media, and source rows that could not be safely resolved are not
included in the image tars.

All included images are:

* fully decoded from audited source bytes and encoded as WebP at quality 95;
* resized only when necessary to a maximum of 4,000,000 pixels while
  preserving aspect ratio;
* never upscaled;
* stored with alpha transparency preserved when present.

## Dataset Structure and Format

Images are distributed across 1,000 zero-padded buckets named `0000` through
`0999`. The bucket for an image is determined by its Gelbooru post ID modulo
1,000:

```bash
BUCKET=$(printf "%04d" $((ID % 1000)))
```

The tar shard and member path are:

```text
original/data-$BUCKET.tar
$BUCKET/$ID.webp
```

For example, image ID `210001` is stored in:

```text
original/data-0001.tar
0001/210001.webp
```

### File Tree

```text
/
├── README.md
├── metadata/
│   └── posts.parquet
└── original/
    ├── data-0000.tar
    ├── data-0001.tar
    └── ... (through data-0999.tar)
```

`metadata/posts.parquet` contains the original metadata fields plus the
release fields `bucket`, `path`, `width`, `height`, `file_size`, `sha256`,
`has_alpha`, `resized`, and `source_file_size`. Each metadata row corresponds
to exactly one released WebP member.

### Minimal Access Example

```python
import io
import tarfile

import requests
from PIL import Image

url = (
    "https://huggingface.co/datasets/nyanko-devs/gelbooru2026/"
    "resolve/main/original/data-0001.tar"
)
response = requests.get(url, timeout=120)
response.raise_for_status()

with tarfile.open(fileobj=io.BytesIO(response.content), mode="r:") as archive:
    member = archive.getmember("0001/210001.webp")
    with archive.extractfile(member) as handle:
        image = Image.open(handle)
        image.load()
```

For bulk training, stream or download shards rather than loading an entire tar
into memory as in this small example.

## Notes

The processed release does not preserve original source encodings. Use the
metadata `sha256` field to verify a released WebP member and `file_size` to
verify its length. Exact cross-dataset deduplication applies to this frozen
release and the frozen Danbooru2026 release used during construction.