YuNet Model Loader

YuNet Model Loader is the most boring node in this pack, and that's the compliment. It has no inputs, no widgets, and exactly one job: hand Image Crop Face (YuNet) a working face detector so you can crop faces out of a photo or a batch. You wire its single yunet_model output into that node's yunet_model input, and you're done. It's the detector-loading half of a two-node face-crop setup.

Why is that worth an article? Because of what it doesn't do. It doesn't download a model the first time you use it. It doesn't need OpenCV, a cascade file, or a path you have to get right. The weights - a yunet.safetensors file - ship inside the pack itself, and the detector runs in torch on whatever device ComfyUI is already using. There is nothing to fetch and nothing to configure. That "weights ship with the pack" line is doing real work: most detector setups in ComfyUI are a download-and-place-your-weights ritual, and this one just works.

Under the hood it's a thin loader: the pack builds the YuNet network once and caches it for the process, so the second run and the fiftieth run cost nothing extra. If the config gate is ever off you'll get an availability error that names the key - features.yunet in the pack's config.yaml, which defaults to on.

Why you'd bother

Face-detection crops are a standard step in two workflows. One: batch-detailing - run detection, crop each face, upscale or detail-pass each crop, paste back. Two: feeding a training or tagging pipeline that wants faces extracted. If you want a single crop from one picture you'd probably rather just do it by hand; the node earns its keep when there are dozens of faces or you want them handled consistently, every run.

Install

It's part of WAS Node Suite v3. ComfyUI Manager → search "WAS Node Suite v3", or:

cd ComfyUI/custom_nodes
git clone https://github.com/WASasquatch/was-node-suite-comfyui

Restart. Requires ComfyUI 0.14.0+ and Python 3.10+, and - again, the selling point - nothing gets installed or downloaded at setup because the weights rode in with the pack. If an old workflow loads and complains the node is missing, it's the usual fix: make sure the pack updated and restart.

The only real gotcha is temperamental: if you loaded Image Crop Face (YuNet) and skipped this loader, the crop node has no detector and will tell you so. The wire order is loader → crop, and the loader's output only fits that crop node's input. Short, honest, and once it's wired you can forget it exists - which is exactly what a loader should be.