Animadex Character Selector

If you've ever stared at an anime prompt trying to remember whether it's "Hatsune Miku" or "hatsune_miku, vocaloid, ahoge, twintails," this node is for you. Animadex Character Selector drops a searchable dropdown of 36,492 anime characters onto your graph, and picking one hands you three things at once: a reference image, the character's trigger words, and its Danbooru-style feature tags. No LoRA hunting, no tag-soup typos.

The trick that makes it useful: this isn't a random fan list. The bundled animadex_top_characters.json is scraped from animadex.net, the community search engine built by generating ~49,000 sample images to find out exactly which characters and tags the Anima model actually knows, ranked by Danbooru post count as a proxy for how well it learned each one. So the trigger strings it outputs ("hatsune miku, vocaloid") come pre-validated against real generations. That's the part no other character pack offers.

How it works

The node loads the JSON once at import and looks up whatever "Name (Copyright)" you picked - the enum format is like Hatsune Miku (Vocaloid). It then downloads the character's reference image over HTTP from animadex.net's blob storage, converts it to a float IMAGE tensor, and returns the trigger words, the feature tags, and a combined string. There's also a frontend gallery: the node gets an "Open Gallery 🖼️" button that pops a browsable modal with search, copyright filter, random, and favorites (saved to localStorage), backed by API routes the pack registers on ComfyUI's own server. No separate web app, no account.

The inputs and outputs that matter

Only two inputs, and you'll touch one of them 99% of the time:

• character - the searchable dropdown. Type to filter; it's 36k entries, so the widget is chunky but the filter makes it fast.
• use_high_res_image - false by default. The thumbnail downloads fast and is plenty for a reference image; true pulls the full original and is noticeably slower.

The four outputs are where the value lives:

• image (IMAGE) - the reference picture, ready to wire into IPAdapter or ControlNet for character-consistent generations.
• trigger_words (STRING) - the tag string the character responds to; goes into CLIP Text Encode.
• prompt_tags (STRING) - Danbooru-style feature tags (hair, eyes, outfit details) to enrich the prompt.
• trigger_and_tags (STRING) - both combined, for when you just want one wire.

Installing it

It's a plain custom node with zero extra dependencies - no requirements.txt, and the imports (torch, numpy, Pillow, aiohttp) are all things ComfyUI already ships. No model downloads either.

cd ComfyUI/custom_nodes
git clone https://github.com/wrt122311/ComfyUI-Animadex-Node.git

Restart ComfyUI, then right-click → Add Node → Animadex → Animadex Character Selector. It should also show up in ComfyUI Manager's search if you'd rather install it there.

Gotchas

The big one: it downloads the reference image over the network every time you run the queue - no local caching in the node. Keep internet on, and if a download fails, the code silently returns a blank 64×64 black image and empty strings. That's the failure mode that'll trip you up: your IPAdapter happily processes a black reference and you'll wonder why your character looks like a silhouette.

One more thing to know before you over-trust it: the dataset is validated against Anima. The tags are Danbooru vocabulary, so they read fine on Illustrious, NoobAI, or Pony - but the "the model definitely knows this character" guarantee is Anima-specific. And AnimaDex's index cuts off at December 2024 while Anima's own knowledge runs to October 2025, so a very new character may be missing even though the model knows it. Misses aren't proof of absence; a quick Danbooru check settles it.