anime-art-curated

A curated WebDataset of anime / digital illustration images with Danbooru-style text tags, intended as a clean training corpus for text-to-image and illustration-style model fine-tuning.

Provenance

This dataset is a filtered re-publication of an earlier dataset (advokat/artist390k, now deleted) which was itself algorithmically curated from broader booru sources by selecting popular images. The original source inadvertently contained material that the maintainer chose not to publish.

This filtered version applies a deny-tag filter to remove rows whose human-applied Danbooru tags include any of the following, and was published to replace the original.

Filter applied

Rows are excluded if any of these tags appear in the comma-separated txt field (case-insensitive, underscores normalized to spaces):

Tag	Rejected count
loli	1,427
age difference	1,019
child	324
shota	176
toddlercon	13

Rejected rows are dropped; image bytes for kept rows are forwarded byte-for-byte with no re-encoding. Captions are forwarded verbatim.

The filter relies on Danbooru's human moderation of the loli tag, which is enforced more strictly than most tags due to payment-processor compliance requirements. The filter is therefore reliable for tagged content but cannot catch rows where the same content was untagged or mis-tagged.

Stats

• Source rows: ~391,000 (from advokat/artist390k, 391 shards)
• Kept rows: 388,000 (99.25% retention)
• Total size on HF: ~641 GB
• Shards: 391 WebDataset .tar files (shard-00000.tar … shard-00390.tar)
• Per-shard size: ~1.6 GB average

Format

WebDataset: each shard is a .tar file containing samples grouped by key, with two extensions per sample:

• <key>.png — the image bytes (PNG)
• <key>.txt — comma-separated Danbooru-style tags (single line)

Tag style:

• Lowercase
• Spaces between words within multi-word tags (e.g. blue archive, long hair) — not underscores
• Common tag categories include: character, series, body type, clothing, composition, action, setting, art style, quality

Example caption:

mika (blue archive), 1girl, solo, halo, long hair, pink hair, yellow eyes,
school uniform, smile, sitting, indoors, looking at viewer, ...

Loading

With Hugging Face datasets:

from datasets import load_dataset
ds = load_dataset("advokat/anime-art-curated", split="train", streaming=True)
for row in ds:
    img_bytes = row["png"]
    caption  = row["txt"]
    key      = row["__key__"]
    ...

With WebDataset directly (for high-throughput pipelines):

import webdataset as wds
url = "pipe:curl -s -L https://huggingface.co/datasets/advokat/anime-art-curated/resolve/main/shard-{00000..00390}.tar"
ds = wds.WebDataset(url).decode("pil").to_tuple("png", "txt")

Intended use

• Fine-tuning text-to-image models on anime / digital illustration style
• Training tag-conditioned generative models that consume booru-style tags
• Creative model research where a curated, large-scale anime art corpus is needed

Out-of-scope use

• Photorealistic or non-illustrative image generation
• Identifying specific real people (the dataset contains anime / fictional character art, not photographs)

Limitations and caveats

The deny-tag filter is only as good as the tagging on the source. While the loli tag is one of the more reliably moderated booru tags, no automated filter is perfect; downstream users who require stricter content guarantees should layer additional filters (e.g. image-side classifiers, additional tag exclusions, or human review).

Adult / nudity-tagged content of clearly-adult characters is not filtered out — the deny list targets only the categories listed above. Downstream users who need a fully SFW dataset should add additional deny rules.

License

The original images are from booru-style image board sources. Tagging is crowdsourced human work from those sources. Users of this dataset are responsible for verifying licensing terms applicable to their use case. This re-publication is provided as-is without warranty.

Generation pipeline

This dataset was produced using the streaming filter script in the Anima fine-tune pipeline kit (reference implementation in 08_filter_and_republish.py). The script streams the source WebDataset shards over HTTP, applies the deny-tag filter, and re-publishes only passing rows as new WebDataset shards.