ComfyUI-Generator-Pack
Danbooru-based tags generator and conflict filter
ComfyUI-Generator-Pack
Suggests the tags most relevant to your prompt from a 5.48M-post Danbooru dataset, and drops the ones that contradict it.
Usage
Feed a prompt into Tags Generator, read the extended prompt out of processed_text.
| Widget | What it does |
|--------|--------------|
| n | How many tags to add, counted after post-processing. 0 = draw the length from the corpus and stop once nothing beats chance |
| characters / pose / expressions / body / clothes / background | A toggle and a _share each. The share is relative to the categories still on: with only pose 0.2 and expressions 0.1, ten tags come back 7 and 3. -1 = no cap, 0 = off |
The shares are what keep one axis from taking the draw over, since every pick conditions the next. characters
owns the subject itself (1girl, solo), so it anchors the gender of everything after it; background
carries objects and compositions on its one share.
| Reach for | When |
|-----------|------|
| rating | The explicitness is off. all draws a tier from the seed, so a new seed rerolls it |
| momentum | The tags have nothing to do with each other (raise), or read as one runaway scene (lower) |
| repetition_penalty | Half the draw goes on respelling one idea — blue skin, pale skin, dark skin |
| lift_threshold | The output contradicts the prompt in ways the data merely discourages. 0.1 catches only pairs that essentially never co-occur |
| blacklist | A regex over the candidates (hair\|eyes, ^black ). <color> expands from resources/wildcards.yaml. Filters candidates, so n still holds |
| temperature / top_k / top_p / min_p / seed | The usual sampling knobs. temperature 0 is argmax |
| filter_copyright | On by default: drops candidates owned by one character or franchise — the tags that turn a library into Patchouli's library. resources/copyright_blacklist.txt adds the ones the statistics miss, and is meant to be edited. Your own tags are never dropped |
| order_tags | On by default: the added tags come back grouped by kind — subject, body, expressions, pose, clothes, scene. Off keeps the draw order. The input prompt is never reordered |
Every widget carries its own tooltip — hover it with the node in front of you.
Three things happen without a widget for them. An empty prompt draws an anchor the corpus has plenty to say
about (beach, moon) and grows a scene from it. A weight at or below zero reverses a tag instead of
asserting it: (light particles:-1.2) costs light particles and what travels with them. A one-person prompt
vetoes the tags that need a second character — the list is resources/solo_conflict.txt, on disk after the
first run, and it is meant to be edited: your copy is never overwritten.
Example
workflows/comfyui-generator-pack-workflow.json
Installation
Search for ComfyUI-Generator-Pack in ComfyUI Manager, or:
cd ComfyUI/custom_nodes
git clone https://github.com/alchemine/comfyui-generator-pack
Nodes (GeneratorPack/Tags)
Tags Generator — the sampler above.
Tags Conflict Filter — drops tags contradicting the fixed ones, by co-occurrence lift rather than a hand-written conflict list. Tags Generator's veto as a node of its own, for tags from anywhere else.
Classify Tags — splits a prompt into characters, clothes, body, expression, pose, background, objects, nsfw, others.
Group Tags — one themed group per line; tags sharing a last (or first) word gather together, person tags
lead. cap trims each group, colours first.
Data
Nothing ships in the repository. Every table, label file and list downloads into resources/ the first time a
node needs it, each pinned by sha256 — the small files as one archive, the statistics tables (up to 100MB)
separately, so a workflow that only groups tags never pulls what the sampler needs. A file already on disk is
never re-fetched, so an edited solo_conflict.txt, copyright_blacklist.txt or wildcards.yaml stays yours.
License
GPL-3.0 — see LICENSE.