IPAdapter Unified Loader FaceID V2

FaceID is the variant of IP-Adapter that swaps the CLIP image embedding for an InsightFace ArcFace face-recognition vector, so instead of "make it look generally like this image," it means "make it look like this person." It's the no-training face tool that defined the SD 1.5 era - and it's fiddlier to set up than plain IP-Adapter because it needs a face-encoding model and a companion LoRA on top of the usual CLIP vision and adapter files. This loader is the version that wires all of that for you from a single preset dropdown.

Like the standard IPAdapterUnifiedLoaderV2, it inspects whether your base model is SD 1.5 or SDXL, then picks the adapter file, the CLIP vision encoder, the ArcFace model, and - where needed - the companion LoRA by name. What it adds is the stuff FaceID specifically needs:

• lora_strength (default 0.6) - the weight applied to the FaceID companion LoRA. The community runs it anywhere from 0.5 to 1.0; it's the second knob of identity strength, distinct from the adapter weight you set later.
• provider - CPU, CUDA, ROCM, DirectML, OpenVINO, CoreML. This selects which backend InsightFace runs on. If you're on Windows with an AMD card you'll want DirectML; on Linux ROCm boxes, ROCM. CPU works but is slow.

Presets: FACEID (base), FACEID PLUS - SD1.5 only, FACEID PLUS V2 (the variant most workflows settled on), FACEID PORTRAIT (style transfer), and FACEID PORTRAIT UNNORM - SDXL only (strong). The "SD1.5 only" and "SDXL only" tags are enforced - pick the wrong one for your base and you get a clean error.

Outputs are the same shape as the standard loader: a MODEL with the LoRA merged in, and an ipadapter pipe carrying the CLIP vision model, adapter, and InsightFace handle. That pipe goes into a FaceID-capable apply node (the Advanced node detects the face via InsightFace and crops it before encoding).

The licensing trap, up front

FaceID reads ArcFace embeddings from InsightFace, whose pretrained models are research-only. The model card states it plainly: IP-Adapter-FaceID is "released exclusively for research purposes and is not intended for commercial use." The plain CLIP-based IP-Adapter variants are Apache 2.0 and sellable; anything with FaceID in the name is not. If you're making money off the output, use the PLUS FACE preset on the standard loader instead - it captures a face without face recognition and stays commercial-safe.

Install

Pack: chflame163/ComfyUI_IPAdapter_plus_V2 (V2 fork of cubiq's, GPL-3.0). Via Manager or:

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

The one install step that isn't automatic: InsightFace is not a dependency of this pack - the loader imports it lazily and fails at runtime if it's missing. Install it into your ComfyUI environment:

pip install insightface

You also need the InsightFace models (buffalo_l) in ComfyUI/models/insightface/, the FaceID adapter + LoRA in models/ipadapter/ and models/loras/, and the CLIP vision encoder in models/clip_vision/.

Where people get burned

InsightFace install is the classic headache - it needs a working onnxruntime and compiling insightface itself can trip on Visual C++ / cmake depending on your platform. If the loader throws an ImportError, that's the cause, not anything in this pack. Second: provider mismatch. Choosing CUDA without a working CUDA onnxruntime produces cryptic failures; CPU always works for testing. And the licensing caveat above is the one people hit months later, when they try to sell something built on a FaceID preset.