Imgutils DeepDanbooru Tagger

DeepDanbooru is the granddaddy of danbooru taggers - a small, fast model that's been around long enough to feel like infrastructure. Imgutils DeepDanbooru Tagger wraps it with two threshold knobs and a drop-overlap toggle, and returns rating, general, and character tags. It's the lightest option in the pack's tagger lineup, which is both its selling point and its ceiling.

Why you'd reach for it

Speed and footprint. The taggers in this pack form a ladder - DeepDanbooru at the bottom (smallest, fastest, oldest), DeepGelbooru in the middle, Camie at the top (huge vocab, heavier). If you're tagging thousands of images and don't need last-word accuracy, or you're running on a machine where every model download counts, DeepDanbooru's lightness is a real feature. It's also the classic "is this roughly what's in the image" tagger for quick checks.

How it works

It wraps get_deepdanbooru_tags from imgutils. The inputs mirror the other sectioned taggers:

• general_threshold (FLOAT, default 0.5, 0–1, step 0.05) - confidence cutoff for general tags.
• character_threshold (FLOAT, default 0.5, 0–1, step 0.05) - cutoff for character tags.
• drop_overlap (BOOLEAN, default false) - remove redundant/overlapping tags.

Outputs: tags (STRING, comma-separated) and json (STRING, sectioned into rating/general/character).

Honest take

Let's be straight about the model's age: DeepDanbooru is the old guard, and it shows. Its tag vocabulary is smaller and its recognition is rougher than the newer taggers - on modern anime it will miss tags that DeepGelbooru or WD14 catch. Its defaults are also stricter (0.5/0.5 vs DeepGelbooru's 0.3/0.3), so out of the box you get a leaner, surer list. If accuracy is the goal, DeepGelbooru is a near-straight upgrade with barely more cost, and Camie wins on vocabulary. Choose this one deliberately: you want the small, fast model, or you're matching an old pipeline. Otherwise you're leaving tags on the table.

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 DeepDanbooru model is small and downloads on first use to ~/.cache/huggingface/hub/.

Troubleshooting

Tags coming out too sparse for your taste? Drop both thresholds toward 0.3 and re-run - DeepDanbooru is conservative by default. Output cluttered with near-duplicates? drop_overlap on. And if the results feel dated compared to what you see elsewhere, that's the model, not the node - the fix is a newer tagger in the same pack, not tuning.