Load WD14 Tagger

If you've ever dropped a generated image into a booru tagger to get its prompt back, you already know what this node does - WD14 is the standard Danbooru-vocabulary tagger, the same family the captioning tools use. Load WD14 Tagger is SimpleSyrup's version: a model loader that produces a WD14_TAGGER socket consumed by the pack's tagging nodes (Tag SEGS w/ WD14 and Tile & Tag SEGS). It exists because those nodes need a real, loaded tagger to chew on, and this is the clean way to give them one.

Why WD14 at all

The tags are the point. WD14 (SmilingWolf's ONNX conversion of the WD14 model family) emits Danbooru-style tags - 1girl, long hair, standing, the whole vocabulary. If your checkpoint is a Danbooru-trained model like Illustrious or NoobAI, that vocabulary is the model's native language, so tagging a region and feeding the tags back as a prompt is a genuinely strong loop. It's also the reason the tagger matters for the pack's detailer story: when Tile & Tag SEGS cuts an image into regions, it needs something to describe each region, and WD14 is that something.

How it works

The dropdown lists the pack's catalog of known SmilingWolf WD14 models. The default, and the one most people should start with, is wd-eva02-large-tagger-v3 - the Eva02 large is the best-accuracy pick in the family. The other entries (vit, swinv2, convnext variants, and the older v1-4 models) trade accuracy for speed or footprint. Select one, and if it isn't on disk the loader pulls the ONNX model and its tag CSV from Hugging Face with visible Comfy progress, then loads it into an ONNX session. The node uses the pack's onnxruntime dependency for this, and the loaded tagger is cached so repeat runs don't reload the weights.

The single wd14_model input is the only thing you set, and the single wd14_tagger output feeds the tagging nodes. Nothing else wires in - no image, no threshold; that's the consumer nodes' job.

Common issues

First load means a download, and the Eva02 model is the biggest of the lot - on a slow connection, budget a few hundred MB and a wait. There's also a model-folder gotcha: this loader looks in SimpleSyrup's own registered model folders, so a WD14 ONNX you fetched for another tool won't be auto-found here. And the output is a pack-specific socket type, so it won't plug into tagger nodes from other packs - if you're mixing ecosystems, check which tagger a workflow actually wants.

The bigger conceptual gotcha is vocabulary matching, same as any WD14 tagger: the tags are Danbooru syntax, and they pay off on Danbooru-trained models. On a natural-language model like Flux, feeding it 1girl, long_hair, standing is a different prompt paradigm entirely - useful only if you know your model speaks tags.

Installing it

It ships in the SimpleSyrup pack. ComfyUI Manager: search SimpleSyrup, install, restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Artificial-Sweetener/SimpleSyrup.git
cd SimpleSyrup
../venv/bin/python -m pip install -r requirements.txt

That install brings in onnxruntime (a chunky dependency) plus the rest of the pack's requirements, so the first install is heavier than a lone tagger would be. You don't need to install SmilingWolf's tagger repo separately - this loader replaces it.