Imgutils DeepGelbooru Tagger

If you want danbooru tags and you don't need Camie's enormous vocabulary, DeepGelbooru is the sweet spot: noticeably better than DeepDanbooru at recognizing what's in the image, without the weight of the big taggers. Imgutils DeepGelbooru Tagger wraps the model with the same two-threshold-plus-overlap interface as its sibling, and returns rating, general, and character tags.

Why you'd reach for it

It's the accuracy middle ground in the pack's tagger ladder. The KB's captioning docs keep pointing anime-LoRA trainers at danbooru taggers because tags match how Illustrious/NoobAI/Pony-adjacent bases were trained; DeepGelbooru is the sensible default when you want better-than-DeepDanbooru tagging without Camie's model size and first-run download. It's the "just tag this folder" node.

How it works

It wraps get_deepgelbooru_tags from imgutils. Same shape as the other sectioned taggers:

• general_threshold (FLOAT, default 0.3, 0–1, step 0.05)
• character_threshold (FLOAT, default 0.3, 0–1, step 0.05)
• drop_overlap (BOOLEAN, default false)

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

Honest take

Two things to know before you run it. First, the defaults are low - 0.3/0.3 - which means out of the box you get more tags than DeepDanbooru, including more noise. For prompts that's often fine; for dataset captions you may want to raise them or lean on drop_overlap to keep captions clean. Second, DeepGelbooru's training set is newer and its vocabulary wider than DeepDanbooru's, so it catches tags the old model simply never learned - for most people this is the "correct" danbooru tagger until they hit a tag Camie knows and this one doesn't. And on that note: this model is still a notch below WD14 or Camie on modern-content recognition, so if you're chasing the long tail, that's the upgrade path, not more tuning here.

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]. Model downloads on first use to ~/.cache/huggingface/hub/ - modest size, quick enough.

Troubleshooting

Captions feeling cluttered? That's the 0.3 defaults working as designed - raise the thresholds or turn drop_overlap on. Tags missing on clearly-modern anime? The model's vocab has limits; that's the Camie/WD14 call. And if you chain this with another tagger, feed both into Imgutils Deduplicate Tags to merge the overlap into one clean list.