Danbooru

Writing a booru-tag prompt by hand for an anime checkpoint is basically guessing at a vocabulary the model was trained on, one tag at a time, hoping you didn't skip the one that mattered. The Danbooru node in Lazy Pony Prompter skips the guessing entirely: point it at a Danbooru search, and it pulls the actual tags real artists put on real images, then reshapes them into a prompt formatted for your checkpoint. You're not writing a prompt from scratch - you're borrowing one a human already proved works.

Despite the pack's name, this node isn't Pony-only. LPP started out pulling from Derpibooru and e621, and Danbooru was added in the 1.1.0 release to widen the net to anime-tag checkpoints - NoobAI, generic Danbooru-trained anime models, and Pony itself if you'd rather source from anime art than furry art. Danbooru's tag vocabulary is, per the KB's own writeup on it, "the prompting language of the whole SDXL anime ecosystem" - so if your checkpoint understands booru tags at all, this node speaks its native language.

How it works

It's a fetch-and-reshape pipeline. You hand it a search query - same syntax as typing into Danbooru's own search box - and it pulls tag metadata from count matching images. From that pool it prunes tags nobody wants in a prompt (pure meta noise), pushes the important stuff like character names toward the front, and randomly samples one resulting prompt to hand you. Run it again and you get a different random pick from the pool, or a fresh pool entirely if you re-query.

The inputs and outputs that matter

• query - your Danbooru search string. Character name, series, pose tag, whatever narrows the pool down.
• tag_filter - comma-separated patterns to strip or swap out of the pulled tags before they hit your prompt, using the same glob syntax as the README's Quick Filter (* for anything, ? for one character, horn||wings to substitute one tag for another). Handy for a one-off "get rid of this" without touching the query itself.
• count (5–1500, default 100) - how many images' worth of tags to pull from. More gives you more variety to sample from; no reason to max it out just testing a query.
• send_request - the field to actually watch. Off by default, and it's ComfyUI's stand-in for the "Send" button the A1111 version has: flip it true to fire the network query. Leave it off and the node works from whatever tag_data you've wired in instead of hitting the API every single queue.
• rating and sort_type - append a content-rating tag (General/Sensitive/Questionable/Explicit) and a sort order (Score, Faves, Rank, and friends) straight onto your query.
• format - decides how pulled tags assemble into a prompt string: plain Generic Anime, a variant that keeps underscores, or profiles tuned for NoobAI XL v1.0 and Pony Diffusion V5(.5)/V6. Match this to your checkpoint, not to the booru you queried - pick Pony and the output opens with Pony's mandatory score_9, score_8_up, score_7_up... string, because forgetting that string is the single most common reason a Pony-family gen comes out wrong.
• Optional prompt_template overrides the default layout using LPP's own tokens ({character}, {species}, {rating}, {artist}, {general}, {meta}, {prompt}).
• Optional tag_data (type LPP_TAG_DATA_DANBOORU) accepts a previously loaded collection - see the Tag Data Loader node - so you can re-roll prompts from a cached pull without querying again.

Outputs: Prompt (STRING) wires straight into your CLIP Text Encode. LPP Tag Data (the generic LPP_TAG_DATA type) is the raw pulled collection - send it to the Tag Data Saver node if this pull is worth keeping around under a name instead of losing it on the next run.

How to install it

Easiest path is ComfyUI Manager: search "Lazy Pony Prompter," install, restart. Manual install is the usual two steps:

cd ComfyUI/custom_nodes
git clone https://github.com/Siberpone/lazy-pony-prompter

Restart ComfyUI afterward. No heavyweight dependency chain and no model to download - it's a thin API client, not a checkpoint. What you do need is an active internet connection at generation time, since it's calling Danbooru's live API. The author also ships a demo image (a Princess Luna picture - drag it onto your canvas) preloaded with a basic workflow and notes on every LPP node, the fastest way to see how the pieces wire together.

Common issues & troubleshooting

Nothing happens when you queue. Check send_request first - it defaults to off, and if you forgot to flip it and haven't wired in cached tag_data either, the node has nothing to work from.

Empty or near-empty results. A specific query stacked with a narrow rating filter can return zero matching images, and the node has nothing to build a prompt from. Loosen the query or the rating before assuming something's broken.

Output doesn't match what your checkpoint expects. Almost always a format mismatch - you fetched Danbooru tags but picked the wrong prompt profile for your model. The format should match the checkpoint, always.

Things feel a little rough. The author says so directly in the README: the ComfyUI nodes are "somewhat experimental and may have some bugs," with "UX compromises" made for ComfyUI's interface limits. If it's clunky rather than broken, that's expected - not a sign you set it up wrong.