IPAdapter InsightFace Loader V2

IPAdapter InsightFace Loader V2 loads the face-analysis model that every FaceID node needs. It's a two-settings loader with one output, and it's simultaneously the most boring node in the pack and the one that causes the most new-user pain, because everything interesting about it - the missing Python package, the model files that don't auto-download, the provider choice - happens around it rather than inside it.

What it does

Two inputs. provider is CPU, CUDA, or ROCM - which backend InsightFace uses for face detection and embedding. CUDA is what you want if you have an NVIDIA GPU; CPU works but is slow and genuinely fine for a single face reference. model_name is buffalo_l or antelopev2. buffalo_l is the default and the only pack InsightFace auto-downloads - the pragmatic choice. antelopev2 is more accurate (it's the one InstantID expects) but must be fetched by hand and placed under ComfyUI/models/insightface. The single output is INSIGHTFACE, which wires into the insightface input on the FaceID apply nodes.

The gotcha is the install

Here's the thing the loader can't do for you: it imports insightface at load time, and if that Python package isn't installed, the node just raises an error. That install has a reputation. For years there was no compatible wheel for newer Python versions, which meant compiling from source - on Windows, Visual Studio build tools. InsightFace 1.0 (May 2026) finally removed the C++ build requirement, so on a modern setup it's usually just:

pip install insightface

Then make sure the model pack is where the loader looks: ComfyUI/models/insightface. buffalo_l downloads itself on first use; if you want antelopev2, fetch the zip from its hosted link and unzip it into that folder manually.

When you'd use it

Anytime you're on the FaceID path and aren't using a unified loader. The IPAdapter Unified Loader FaceID V2 node bundles insightface loading in with everything else, so this standalone loader is for the explicit pipeline: IPAdapter Model Loader V2 + this + a CLIP vision loader, feeding a FaceID apply node directly. It's also the way to load one insightface model once and reuse it across multiple FaceID nodes without reloading.

Installing it

The pack itself installs like everything else here:

cd ComfyUI/custom_nodes
git clone https://github.com/chflame163/ComfyUI_IPAdapter_plus_V2

Restart, then handle the insightface side as above. And remember the license wall: InsightFace's weights are non-commercial, and anything built on them - every FaceID output - inherits that. This pack is a renamed snapshot of cubiq's V2 codebase, so no auto-updates from upstream.