ComfyUI-WD-Timm-Tagger

ComfyUI custom node for image tagging using WD v3 models, based on timm.

Installation

1. Clone this repo into ComfyUI/custom_nodes/
2. Install dependencies:

   pip install -r requirements.txt
   

Performance

Tested on 291 images with RTX 5090, wd-eva02-large-tagger-v3 model.

Node	Time
ComfyUI-WD14-Tagger	226.99s
ComfyUI-WD-Timm-Tagger	9.34s

Models

Models are automatically downloaded to models/wd_taggers/ on first use.

Model	Original Repo
eva02	SmilingWolf/wd-eva02-large-tagger-v3
vit-large	SmilingWolf/wd-vit-large-tagger-v3
vit	SmilingWolf/wd-vit-tagger-v3
swinv2	SmilingWolf/wd-swinv2-tagger-v3
convnext	SmilingWolf/wd-convnext-tagger-v3
pixai-tagger-v1.0	pixai-labs/pixai-tagger-v1.0
pixai-tagger-v0.9	pixai-labs/pixai-tagger-v0.9
OppaiOracle-V1.1	Grio43/OppaiOracle

You can also place your own model folders under models/wd_taggers/ and they will appear in the model list automatically.

Inputs

Name	Type	Description
image	IMAGE	Input image(s)
model_name	COMBO	Model to use
dtype	COMBO	Model precision (auto / bf16 / fp16)
general_threshold	FLOAT	Confidence threshold for general tags (default: 0.35)
character_threshold	FLOAT	Confidence threshold for character tags (default: 0.75)
add_rating	BOOLEAN	Prepend rating tag to output
exclude_tags	STRING	Comma-separated tags to exclude
batch_size	INT	Number of images per inference batch

batch_size only takes effect when multiple images are passed as a batch.

Outputs

• STRING — Formatted tag string per image
• RAW — Dict with ratings, character, general scores per image

Credits

Based on wdv3-timm by neggles.