Cyberdelia Danbooru Prompt

Illustrious, NoobAI, Pony - the whole SDXL anime family was trained on Danbooru's tag vocabulary, and those models respond to a precise tag like long_hair or looking_at_viewer far better than to a sentence about it. The problem is that writing real Danbooru tags means either living in the wiki or trusting an LLM to invent them. And LLMs don't invent real tags. They confidently emit tag-shaped phrases that were never trained into the checkpoint, which is how you get a "good" prompt that quietly underdelivers.

Cyberdelia Danbooru Prompt is the fix. You type a plain-language scene description, it runs through a local LLM, and then - this is the whole point - every candidate is validated against a bundled ~140k-tag Danbooru vocabulary before it ever reaches your text encoder. It's one of three nodes in the cyberdeliaAI/comfyui-cyberdelia-z-engineer pack, and it's the one you reach for when you want booru-quality tags without the manual proofreading.

How it works

The mode input picks the pipeline, and the three modes are the whole story:

• engineered (LLM) - your description goes to the LLM, which proposes candidate tags. Those get validated, formatted, and wrapped in the template.
• validate tags - no LLM call at all. Paste an existing comma-separated tag prompt and it validates, sorts, and reformats it.
• raw positive - exact passthrough. Nothing touches it, not even validation. Use this when the prompt must stay byte-for-byte.

That validation step is where the value lives. Each comma-separated item is checked as one complete tag, so an invented phrase can't survive as an unrelated single-word fragment. It also won't call home: the vocabulary is a static snapshot shipped with the pack, not a live connection to Danbooru. Nothing phones out for tag data, which in a category that once shipped malware dressed as an "LLM vision node" is a quietly good thing.

Inputs and outputs that matter

The defaults are sane, so you really set a handful:

• mode - engineered (LLM), validate tags, or raw positive.
• prompt - your natural-language description, or the existing tag prompt in validate mode.
• prompt_template - the wrapper, default {prompt}. Quality anchors go here: masterpiece, best_quality, {prompt} for Illustrious, or Pony's score_9, score_8_up. They're deliberately not in the LLM's job, so they don't get filtered, reordered, or doubled.
• api_url - defaults to http://localhost:1234/v1 (LM Studio). Point it at any OpenAI-compatible server.
• tag_format - spaces (blue eyes) or underscores (blue_eyes). Spaces is the safer default.

Two optional knobs are worth knowing. fuzzy_cutoff is off at 0 because fuzzy matches change meaning; if you turn it on, 0.85 or higher is the safe start. min_post_count filters rare tags, and anything the LLM proposed but validation rejected lands in dropped_tags - read it, it tells you exactly what got eaten.

Outputs: prompt (validated tags wrapped in your template), tags (just the tags), dropped_tags (the rejects).

Installing it

The pack needs only requests and a running OpenAI-compatible chat-completions endpoint - no model downloads, because the LLM lives in LM Studio/Ollama and the vocabulary ships in the pack.

cd ComfyUI/custom_nodes
git clone https://github.com/cyberdeliaAI/comfyui-cyberdelia-z-engineer.git
pip install -r comfyui-cyberdelia-z-engineer/requirements.txt

Restart ComfyUI, or just search Cyberdelia in ComfyUI Manager and let it handle all of that.

Where people get burned

The LLM connection only happens in engineered mode - if LM Studio isn't running, the default error_mode: stop kills your whole workflow with an error on the node. Set it to fallback_input or empty if you'd rather not block. Also: the default system prompt is tuned for Illustrious. If you're on Pony, keep score_9 in the template, because a tag with no Danbooru home gets dropped to dropped_tags like everything else. That output list is your friend - when the result looks thinner than your idea, check what validation threw away before you blame the checkpoint.