Methodology & provenance
How Latent Tags builds its research
Where the tag data, model previews and trends come from — and what each signal can and cannot tell you.
931,802 tags · updated Aug 16, 2026
Tag vocabulary
Canonical names, categories, aliases, implications and wiki groups come from Danbooru. An offline ETL pipeline normalizes them into a read-only snapshot; the site never edits Danbooru or fetches its API during a visit.
Model evidence
Preview images are resolved from artwork already in the Latent corpus. A missing preview means the corpus has no matching example yet — it does not prove that a model cannot understand or draw the tag.
Prompt output
The workbench preserves canonical tag categories, then applies each model's documented artist naming and weight conventions. Runtime syntax such as A1111 or ComfyUI is kept distinct from the model family itself.
Trend counts
Charts show monthly Danbooru post counts, not search volume or model popularity. The current upstream series ends in April 2025, so the interface labels it as a frozen snapshot.
Sources and attribution
- DanbooruCanonical tags, categories, aliases, implications, wiki groups, excerpts and related-tag signals.
- deepghs/site_tagsBulk tag seed; CC BY 4.0.
- ffdkj translation tableSimplified Chinese translations; the upstream repository currently declares no license.
- p1atdev/danbooru-ja-tag-pair-20241015Japanese translation fallback; CC0 1.0.
- dataproc5 monthly tag countsMonthly tag counts through April 2025; the upstream dataset currently declares no license.
Latent Tags is an independent research tool and is not affiliated with Danbooru or the creators of NovelAI, Anima, NoobAI, or Illustrious.