InstantID Face Analysis

Every InstantID workflow starts here, and most InstantID problems trace back here too. InstantID Face Analysis loads InsightFace - the face-recognition library that turns a photo into an identity vector - and hands that loaded engine downstream to the nodes that actually do something with it. It doesn't touch your image yet. It just boots the analyzer.

That sounds trivial, and the node itself is trivial: one dropdown, one output. But InstantID's entire trick - getting a recognizable face out of a single reference photo with zero training - rests on InsightFace's ArcFace model correctly reading that face first. Get this node wrong and everything downstream (Apply InstantID, the Advanced version, the attention patch) either errors out or quietly produces garbage.

How it works

Under the hood this node initializes InsightFace with the antelopev2 model pack - not the buffalo_l pack that InsightFace normally auto-downloads for you. That distinction matters a lot, more on it below. Once loaded, it detects the face in whatever image you feed the next node, aligns it, and produces the embeddings that carry facial geometry rather than just "looks kind of like this."

The input and output that matter

There's exactly one input: provider, an enum with four choices - CPU, CUDA, ROCM, CoreML. Pick whatever matches your hardware: CUDA for an Nvidia card (the fast path and what almost everyone should use), CPU if you don't have a GPU or you're isolating a bug, CoreML on Apple Silicon, ROCM for AMD. There's nothing else to configure.

The output is a single FACEANALYSIS object, which you plug into the insightface (or faceanalysis) input on Apply InstantID, Apply InstantID Advanced, InstantID Patch Attention, or Face Keypoints Preprocessor. It's the identity-analysis backbone the rest of the pack builds on.

Installing it

This node ships with the pack itself, so grab the whole thing: search "ComfyUI InstantID (Native Support)" in ComfyUI Manager, or cd ComfyUI/custom_nodes && git clone https://github.com/cubiq/ComfyUI_InstantID and restart. The author is Matteo (cubiq), the same person behind ComfyUI_IPAdapter_plus - if you've used IPAdapter you already trust his node quality. Note the repo's been in "maintenance-only" mode since April 2025: it still works, but don't expect new features.

The part that actually causes trouble is what this node depends on, not the node itself. You need insightface, onnxruntime, and onnxruntime-gpu installed as Python packages, and you need the antelopev2 InsightFace model pack downloaded by hand and unzipped into ComfyUI/models/insightface/models/antelopev2 - it is not auto-downloaded like the default buffalo_l pack is. Miss that step and this node has nothing to load.

Common issues

"The InsightFace nodes don't even show up in ComfyUI." This is a real, recurring complaint, and it's almost always a broken clone or a Manager install that didn't fully unpack - people report the pack's Python files (face_analysis.py, insightface_loader.py) simply missing from their custom_nodes folder. Delete the folder and reinstall clean before you chase anything fancier.

InsightFace won't install, especially on Windows. This is the single worst dependency in the whole pack - it needs a compiled wheel matching your exact Python version, and picking the wrong one (or none) is the most common reason this node throws on load. If you're on Windows, grab a prebuilt wheel matching your Python version rather than fighting pip install insightface from source.

Node loads but errors the moment you run it. Almost always the missing antelopev2 model - check that ComfyUI/models/insightface/models/antelopev2 actually has files in it, not just an empty folder from a failed unzip.

Also update ComfyUI itself before installing. The pack has historically required a recent core version; an old ComfyUI install is a quiet source of "it just doesn't work" reports.