YOLO Classifier Model Loader

Why a loader node gets its own article

YOLO Classifier Model Loader is the boring half of the two-node ComfyUI_YOLO_Classifiers pack - and the one you'll hit first, because nothing else in the pack works without it. Its entire job is to produce the YOLO_CLASSIFIER_MODEL handle that YOLO Classify requires. One dropdown, one output, done. It reads as trivial, and it mostly is. But there's one real trap buried in it, and it's the models.

What it does

Pick a model name and the loader resolves it against a folder ComfyUI registers as models/ultralytics/classifiers/, builds an Ultralytics YOLO(...) instance from the weights file, and passes that handle out. The five choices are the stock YOLOv8 classification checkpoints, from tiny to big:

yolov8n-cls.pt - nano, a few MB, fast
yolov8s-cls.pt
yolov8m-cls.pt
yolov8l-cls.pt
yolov8x-cls.pt - extra-large, slowest, up to ~100 MB

For most routing jobs the nano or small is all you need. These are image classifiers, not diffusion checkpoints - a few MB to ~100 MB is the whole range, and bigger buys a bit of ImageNet accuracy that rarely matters for "is this a cat or not."

The one input, the one output

model_name - the only input, an enum dropdown of those five .pt files. If a file isn't in the folder, it won't appear in the list.
Output: "YOLO classifier model", a custom YOLO_CLASSIFIER_MODEL type. Wire it into the yolo_classifier_model input on YOLO Classify. There's nowhere else this type goes - the two nodes are a matched set.

Installing it

Same pack install as YOLO Classify: in ComfyUI Manager, search "YOLO Classifiers", or clone it:

cd ComfyUI/custom_nodes
git clone https://github.com/SuperMasterBlasterLaser/ComfyUI_YOLO_Classifiers

then restart. The pack pins ultralytics==8.2.32, so first launch runs a pip install.

The trap: nothing downloads itself

The dropdown is populated by scanning a folder - there's no "download model" button, and no model manifest for Manager to install either. You place the weights yourself:

mkdir -p ComfyUI/models/ultralytics/classifiers
cd ComfyUI/models/ultralytics/classifiers
wget https://github.com/ultralytics/assets/releases/download/v8.2.0/yolov8n-cls.pt

then restart ComfyUI and the dropdown fills in. The source comment points at the Ultralytics classify docs as the download reference, and all five -cls files live in the same release.

Two ways this bites you. First, if the selected file is missing, the loader doesn't fail gracefully - it hands Ultralytics an empty path, which Ultralytics tries to interpret as its own default model, usually producing a confusing error. Put the file there first. Second, make sure it's the -cls variant. A plain detection yolov8n.pt will load into this node fine, but YOLO Classify reads r.probs, which detection models don't have, so it dies with an attribute error the moment you run it.

One more honesty note: the dropdown is hardcoded to those five names. No custom models, no extra knobs - that's the whole ceiling of this pack. For its one job, it does it cleanly; if you need custom classes or confidence scores, look elsewhere.