ONNXDetectorProvider

This is a loader, plain and simple: point it at an ONNX detection model and it hands you a BBOX_DETECTOR that the rest of Impact Pack can use to find things - faces, hands, whatever the model was trained on. Every detailing workflow needs some detector to say "fix here," and this is the ONNX-format option. In practice most people reach for the YOLO/Ultralytics route these days, so ONNX detection is a bit of a legacy corner - but if you've got an ONNX model you want to use, this is the node that gets it into the graph.

How it works

The detect-crop-refine loop that Impact Pack is built around starts with detection. A detector node produces a BBOX_DETECTOR object, that goes into a Simple Detector or straight into a Detailer, and the pack crops and resamples wherever the detector fired. ONNXDetectorProvider is the front of that chain for ONNX models: it loads the weights and wraps them so downstream nodes see a standard detector, exactly like the BBOX/SAM/CLIPSeg providers do for their formats.

The inputs and outputs that matter

There's really only one input:

• model_name - a dropdown of the ONNX models it found on disk. And here's the catch that confuses everyone: if that dropdown is empty, it's because you have no ONNX detection models installed. The node can't invent one. You need to drop .onnx detector files into the appropriate ComfyUI models folder (the ONNX/bbox detector directory) and restart before anything shows up here.

The single output is BBOX_DETECTOR, which wires into a detailer or a Simple Detector (SEGS) node.

How to install it

Install the Impact Pack, then supply models. Via ComfyUI Manager: search ComfyUI Impact Pack, Install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Impact-Pack comfyui-impact-pack
cd comfyui-impact-pack
pip install -r requirements.txt

then restart. ONNX inference needs the onnxruntime package - the README lists it as an optional dependency, so if the node errors on load or at runtime, install onnxruntime into the same Python environment ComfyUI uses. There's an ONNX tutorial linked from the pack's README worth skimming if you go this route.

Common issues & troubleshooting

The model_name dropdown is empty. No ONNX detector models are installed. Add .onnx model files to your ComfyUI detector-models folder and restart. The node lists what's on disk; it won't download anything for you.

It loads but fails when you actually run detection. Almost always a missing or mismatched onnxruntime. It's an optional Impact Pack dependency, so a default install may not have it. Install it (the GPU build, onnxruntime-gpu, if you want CUDA) into ComfyUI's environment.

Nothing gets detected. Make sure the ONNX model you loaded actually matches your target - a face detector won't find hands. If you're not tied to ONNX specifically, the modern default is the YOLO detector from the separately-installed Impact Subpack (UltralyticsDetectorProvider), which most current tutorials and workflows assume.