ONNX Detection Model Loader

This node doesn't do anything flashy - it loads the two models that everything else in this pack needs, and that's exactly why it's the first node in every Wan Animate workflow that touches the WanAnimatePreprocess pack. Run it, and you've done the hard part.

It comes from kijai, who is effectively the maintainer of "running Wan locally in ComfyUI." His ComfyUI-WanVideoWrapper is the de facto third-party wrapper for the whole Wan family (the KB's Wan guide calls it the primary one), and this preprocess pack is the companion that feeds it. The author's README is a single paragraph, which is typical: the trust is in the code, not the prose.

What it loads

Two ONNX models, both pulled from your ComfyUI/models/detection folder (the pack registers that folder itself, so drop .onnx files there and they'll appear):

• ViTPose - the whole-body pose estimator. It extracts body, face, and hand keypoints from a cropped person.
• YOLO - the person detector that finds the people first, so ViTPose knows where to look.

Both are bundled into one POSEMODEL output, which you feed into Pose and Face Detection or Pose Detection OneToAll Animation. You only ever run this node once per workflow - the heavy lifting (detection + pose estimation) happens downstream.

The three inputs that matter

• vitpose_model - dropdown of .onnx pose models in models/detection. Pick your ViTPose.
• yolo_model - same folder, pick the YOLO detection model.
• onnx_device - CUDAExecutionProvider (default) or CPUExecutionProvider. GPU if you have it; CPU is the "I just want it to run" fallback.

That's it. One output, model, and it's the only thing that comes out.

Install and models

Install through ComfyUI Manager (search ComfyUI-WanAnimatePreprocess) or:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-WanAnimatePreprocess

Restart ComfyUI. The pack's requirements.txt installs onnx, onnxruntime-gpu, and opencv-python - onnxruntime-gpu is the one that can bite on a fresh machine (CUDA version mismatch), but the loader's CPU fallback exists precisely for that case.

Then get the models into ComfyUI/models/detection:

• YOLO: yolov10m.onnx from the Wan 2.2 Animate repo.
• ViTPose: either the Large wholebody ONNX from JunkyByte/easy_ViTPose, or the Huge model from Kijai/vitpose_comfy.

Gotchas that actually bite

The Huge ViTPose is two files. It ships as vitpose_h_wholebody_model.onnx plus a vitpose_h_wholebody_data.bin external-weight file, split that way because of ONNX's file-size limit. Both have to sit in the same directory; if you grab just the .onnx, the loader will fail at session creation because it can't find the weights. Select the .onnx in the dropdown, keep the .bin next to it.

Empty dropdowns. If vitpose_model and yolo_model show nothing, the models aren't in models/detection - the folder only shows up once the pack has run and registered it. Double-check the exact path (ComfyUI/models/detection, not custom_nodes/...).

Manager install flaking. People hit a "failed to find the following comfyregistry" error installing this pack through the Manager; the reliable workaround is to download the repo as a zip and unpack it into custom_nodes manually. It's one of those "works on most machines, fine on yours hopefully" installs.