GR Load InsightFace (PuLID Flux)

GR Load InsightFace (PuLID Flux) is where you pick how your face gets detected and measured. One dropdown - provider with three choices, CPU, CUDA, ROCM - and out comes a FACEANALYSIS object that the apply node uses to find faces in your reference image and turn them into biometric identity embeddings.

This is the "who is this person" half of PuLID, and it's the reason the whole genre of no-training identity tools exists. InsightFace's pipeline - detect the face, align it, run it through ArcFace - produces a vector that encodes facial geometry rather than appearance. It's the same recognition backbone under IP-Adapter FaceID, InstantID, and PuLID, which is why the community occasionally rediscovers with mild alarm that every identity tool is the same face detector wearing different clothes. Here it's a load-bearing dependency, and it's the fiddliest one in the pack.

What happens when you run it

Pick CUDA (or ROCM on AMD, CPU if you're desperate - it works, it's just slow) and the node loads InsightFace's antelopev2 model set into ComfyUI/models/insightface. The first time, that download happens automatically from InsightFace's own servers, so you need internet and a little patience. After that it's local. The output FACEANALYSIS feeds the face_analysis input on GR Apply PuLID Flux.

The usual failure here isn't the download, it's the Python side. insightface is one of the worst installs in local generation, and this pack's requirements.txt pulls in both onnxruntime and onnxruntime-gpu, which is exactly the setup that breaks when pip grabs a too-new onnx version. The community-blessed fix when insightface won't import or errors at runtime:

pip install onnx==1.14.0 onnxruntime-gpu==1.15.0

The "Two devices issue on several passes" error people hit after installing this pack traces back to the same onnx mess, not to the node itself.

The part nobody puts in the workflow screenshot

Antelopev2 - the very model this node loads - is non-commercial. PuLID's own weights are Apache 2.0, and the pack itself carries a permissive license, but the binding restriction is inherited from underneath: the face-embedding models that supply the identity signal aren't free to sell. If this is for a client project, that's the sentence to read twice. There are alternative node packs that offer a FaceNet path precisely to escape it; within this pack, there isn't one.

The rest is the usual

Install is the pack-standard: ComfyUI Manager (search "ComfyUI-PuLID-Flux-GR"), or git clone https://github.com/GraftingRayman/ComfyUI-PuLID-Flux-GR into custom_nodes, then pip install -r requirements.txt (portable: ComfyUI\python_embeded\python.exe -m pip install -r <path>\requirements.txt).

The GR fork also added proper resource cleanup here - cleanup and __del__ handlers that move the model to CPU and empty the CUDA cache when the graph tears down, which was one of the long-standing issues in the original node it forked. So while this is technically the most boring node in the pack - a dropdown and a license footnote - it's also the one most likely to throw at install time. Get it past first run and you'll never think about it again.