Imgutils Camie Tagger

The pack's heavyweight tagger, and it earns the weight: Camie knows 70,000+ danbooru tags. Imgutils Camie Tagger runs the Camie model and returns rating, general, and character tags - which makes it the widest-vocabulary option in the pack's tagging lineup, and the one to reach for when the classic taggers simply don't know the tag you need.

Why you'd reach for it

The KB's danbooru-tags panel and captioning docs keep saying the same thing: for anime bases (Illustrious, NoobAI, Pony-adjacent), your captions should be danbooru tags, because that's the vocabulary the model was trained on. The question is which tagger. WD14 and DeepDanbooru know the classics; Camie's whole pitch is coverage - newer characters, obscure tags, and recent vocabulary that older models never saw. If your character or concept keeps coming back untagged from everything else, Camie is the tagger that probably knows it.

How it works

It wraps get_camie_tags from imgutils. The mode input isn't a raw threshold - it's a preset that picks the tag-selection strategy:

• Balanced (default) - the all-rounder.
• High Precision - fewer tags, higher confidence. The dataset-captioning mode: a wrong tag costs training time, so you want surer tags even if it means a thinner list.
• High Recall - more tags, including noisier low-confidence ones. Good for discovery, bad for clean captions.
• Micro Optimized / Macro Optimized - presets tuned toward those two metrics respectively.

Plus drop_overlap (BOOLEAN, default false) - removes redundant/overlapping tags like a "blue hair" + "blue eyes" style duplicate pair. Outputs: tags (STRING, comma-separated) and json (STRING, sectioned into rating/general/character).

Honest take

You pay for the vocabulary. Camie is the heaviest tagger here - bigger model, real first-run download, slower inference than WD14 or the DeepDanbooru pair. On a plain clean image the difference from WD14 is marginal; where it's worth it is the long tail. Two workflow tips: keep mode on Balanced for prompts, switch to High Precision for dataset captions, and turn drop_overlap on when you're captioning (off if you want raw output). The sectioned json output plays perfectly with the pack's Deduplicate and JSON Filter nodes if you're chaining taggers.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/xiaden/comfyui-imgutils.git
cd comfyui-imgutils
pip install -r requirements.txt

Or via ComfyUI Manager (search "imgutils"). Needs ComfyUI >= 0.25.0 and Python >= 3.10; dependency is dghs-imgutils[gpu]. The Camie model downloads from HuggingFace Hub on first use and caches in ~/.cache/huggingface/hub/ - expect a real download before the first run finishes.

Troubleshooting

First run sits at "executing" for a while? Model download - normal, wait it out. Output so thin you can count the tags? You're probably in High Precision, which is doing its job; switch to Balanced or High Recall. Tags look cluttered or contradictory? Turn drop_overlap on. And if you're chaining multiple taggers and the outputs overlap heavily, run them through Imgutils Deduplicate Tags to merge cleanly.