anima-tagger-artifacts

Pre-built retrieval artefacts for the Anima tag format of the sd-webui-prompt-enhancer Stable Diffusion WebUI extension. Lets the extension's Anima pipeline do real-time embedding-based tag validation and shortlist retrieval without users needing to rebuild a 270k+ entry FAISS index locally.

Contents

File	Size	Description
tags.sqlite	~30 MB	273,025 Danbooru tags (name, category, post count, aliases, wiki). Post-count floor 10.
tags.faiss	~1.1 GB	FAISS FlatIP index of bge-m3 embeddings (1024-dim) for every tag. Artist and character embeddings include co-occurrence signatures (top-12 general tags from their actual Danbooru posts).
cooccurrence.sqlite	~10 MB	Pointwise-mutual-information table for character↔series, character↔artist, series↔character pairs. Enables automatic series-pairing (e.g. hatsune_miku → vocaloid) at query time.
VERSION	<1 kB	JSON manifest with per-file sha256 + size + build date.

Usage

Automatic — the extension's install.py downloads these on Forge startup, verifies sha256 against VERSION, and re-downloads when the upstream hash changes.

Manual (e.g. for other projects):

from huggingface_hub import hf_hub_download

for fname in ("tags.sqlite", "tags.faiss", "cooccurrence.sqlite", "VERSION"):
    hf_hub_download(
        repo_id="freedumb2000/anima-tagger-artifacts",
        filename=fname, repo_type="dataset",
        local_dir="./data",
    )

How it was built

1. Tag metadata + wiki from NSFW-API/DanBooruTagsAndWikiDumpSept2025 (1.59M tags; filtered to post_count ≥ 10 → 273,025 tags).
2. Post-level tag sets from isek-ai/danbooru-tags-2024 (streamed first 500k posts).
3. For each artist/character, compute top-12 co-occurring general tags across their posts → "style signature".
4. Format each tag as "<name> (<category>) | aliases: ... | <wiki excerpt> | associated with: <top co-occurring tags>" and embed with BAAI/bge-m3 (fp16 on GPU, normalized, 1024-dim).
5. FAISS FlatIP index over all 273k vectors.
6. PMI table over the same post sample for character↔series pairing.

Full rebuild: ~10 minutes on a modern GPU.

License

The artefacts themselves are released under CC0-1.0. The upstream Danbooru tag data is public-domain by convention. Refer to the linked source datasets for their own attribution requirements.

The underlying embedding model (BAAI/bge-m3) is MIT-licensed.