Stand-In Processor Loader

This one's the on-ramp. Every other node in this pack - the face cropper, the video preprocessor, the background restorer - takes a FACE_PROCESSOR as an input, and this is the only node that produces one. Nothing downstream runs without it, so it's the first thing you drop on the canvas.

Stand-In is WeChat Vision / Tencent's identity-preserving adapter for Wan video, released August 2025. The headline was how cheap it was to make: Kijai (who wrote ComfyUI-WanVideoWrapper) noted only about 1% of the base model got retrained, and called the results "pretty accurate" if "not quite as versatile as the bigger models such as Phantom." He shipped his own support for it in the wrapper - but the authors felt his implementation of the preprocessing step wasn't quite matching theirs, so they published this repo as the official correction. It says as much right in the README: "Only images processed with this official preprocessor can fully unleash the power of Stand-In." Everything in this pack, starting with this loader, exists to be that correct preprocessing path.

How it works

Under the hood this loads a YOLO face-detection checkpoint - an object detector, not a recognition model. That's a real and slightly unusual choice worth flagging if you've used other identity tools before. InstantID, PuLID, and IP-Adapter FaceID all lean on InsightFace's ArcFace embedding, which actually recognizes a specific face and comes bundled with a non-commercial license and a genuinely painful install history. Stand-In doesn't need that here - the identity signal comes from the adapter model itself during sampling, not from a recognition embedding baked into the preprocessor. All this node needs to do is find a face in a frame and hand back a bounding box so the rest of the pack can crop, mask, and paste around it. That's a detection job, and YOLO is a perfectly sane, lighter tool for it.

The input and output that matter

There's exactly one setting: yolo_model_name (default model.pt) - the filename of your YOLO face-detection weights. There's no bundled download link in the README, so if you don't already have YOLO face weights sitting around, you'll need to track down a face-trained YOLO checkpoint yourself and drop it wherever this node's model folder expects it.

The output is face_processor, type FACE_PROCESSOR - plug it into every other Stand-In node's face_processor input. One loader, reused everywhere.

How to install it

Standard custom-node install, straight from the README:

cd ComfyUI/custom_nodes
git clone https://github.com/WeChatCV/Stand-In_Preprocessor_ComfyUI.git
cd Stand-In_Preprocessor_ComfyUI
pip install -r requirements.txt

Restart ComfyUI after. It should also show up in ComfyUI Manager under its pack title, "Stand-In Official Preprocessor ComfyUI Nodes," if you'd rather search than clone by hand.

Common issues & troubleshooting

Node errors on load or on first run. Almost always the yolo_model_name pointing at weights that don't exist at the expected path. Double-check the filename matches what you actually downloaded and where you put it - this node has no fallback or auto-download behavior described anywhere in the repo.

It's not finding faces reliably. That's not tunable here - this loader just points at the model file. Detection-quality knobs like confidence_threshold live on the downstream nodes (ApplyFaceProcessor, VideoInputPreprocessor, VideoBackgroundRestorer), so if faces are getting missed, go adjust those, not this one.

Remember this is explicitly a temporary node. The README states plainly that KV cache and advanced attention mechanisms aren't implemented yet, and that WeChatCV is working on the real, full-featured official nodes. If something about performance feels rougher than what Kijai described in his own benchmarks (roughly 30 seconds of overhead for an 81-frame 832×480 clip on a 4090 with optimizations on), that's the state of this stopgap pack, not a misconfiguration on your end.