IPAdapterInsightFaceLoader(kolors) - Legacy

Every IP-Adapter FaceID setup needs an InsightFace model to detect and encode faces - and something has to load it. In this pack that something is MZ_IPAdapterInsightFaceLoader. It's a tiny node with one input and one output, and its entire job is to hand you an INSIGHTFACE object you pass to MZ_IPAdapterFaceID. Like the other MZ_IPAdapter* nodes, it's a legacy alias of cubiq's class, superseded by cubiq's own updated pack - but if your workflow references it, it works exactly as it always did.

How it works

The one input, provider, is a dropdown with three options: CPU, CUDA, ROCM. That's it. Under the hood it loads InsightFace's FaceAnalysis app, pointed at the face models in models/insightface/. The provider choice matters because InsightFace can run its ONNX models on different backends - CPU is the safe fallback (slow but always works), CUDA is what you want on an NVIDIA card, ROCM covers AMD. If you're on an AMD card, the ROCM option is a genuine reason to use this node rather than the CPU fallback.

The output, INSIGHTFACE, carries the loaded face-analysis model and feeds straight into MZ_IPAdapterFaceID's optional insightface input. It's also cached - the loader only reinitializes the model when the provider changes, so re-running the graph doesn't re-pay the load cost every time.

Setup

The models are the real install:

# from MonsterMMORPG/tools
antelopev2.zip -> unzip to models/insightface/models/antelopev2/

The folder layout matters: InsightFace expects the .onnx files in models/insightface/models/antelopev2/. And the insightface Python package has to be installed in the environment ComfyUI runs in - this is the single most common failure point for the whole FaceID path, and the error is usually an ImportError or "no model found" that points at this loader. If that happens, pip install insightface and recheck the antelopev2 path before touching anything else.

Install and where it sits

Pack install is the standard one: ComfyUI Manager search "ComfyUI-Kolors-MZ", or git clone https://github.com/MinusZoneAI/ComfyUI-Kolors-MZ into custom_nodes and restart. Full FaceID graph: MZ_IPAdapterInsightFaceLoader → MZ_IPAdapterFaceID (insightface input), with the IPADAPTER object from MZ_IPAdapterModelLoaderKolors carrying the FaceID .bin weights. If you're building this graph from scratch today, cubiq's ComfyUI_IPAdapter_plus (updated) is the supported path - but if you've got the legacy MZ nodes in a saved workflow, this loader drops right in.