ID Embedding Model Loader

Behind every zero-shot identity model sit three machines: something to find a face, something to recognize whose face it is, and something to translate that recognition into a language the diffusion model understands. ID Embedding Model Loader is all three at once. It's the first node in every InfiniteYou workflow and the only one that goes shopping on Hugging Face. Open the pack's example graph and it's the lonely box on the left with three wires coming out, feeding the entire identity pipeline.

Quick context, because you're here for a reason: InfiniteYou (InfU) is ByteDance's zero-shot identity-preservation method for FLUX - the ICCV 2025 paper that people described as "what InstantID did, but on Flux." Same family as PuLID - an ArcFace identity embedding injected into a diffusion transformer - but with a cleaner story on text alignment and aesthetics. This node is the setup cost of that family. It loads three things at once:

• InsightFace AntelopeV2, the face detector,
• ArcFace, the recognition model that turns a face into a 512-dimension identity vector,
• the image projector, a Perceiver-style Resampler (borrowed from OpenFlamingo) that maps that vector into FLUX's 4096-dimension embedding space as a short token sequence.

On first run it quietly downloads everything: image_proj_model.bin into ComfyUI/models/infinite_you, the AntelopeV2 ONNX files into ComfyUI/models/insightface/models/antelopev2, and the ArcFace weights that ride along with the facexlib dependency. None of it is optional and none of it is small - budget a few GB and a patient first run.

The inputs you actually set

Four required inputs, and a beginner touches two of them:

• image_proj_model_name - sim_stage1 or aes_stage2. Those are the paper's two training stages; aes_stage2 is the final, aesthetics-tuned model and the default in every shipped workflow. Start there.
• image_proj_num_tokens - 8 or 16. 8 is the default and works fine. 16 gives the projector more tokens to express identity with, at a real cost. Leave it at 8 until you're chasing likeness.
• face_analysis_provider - CUDA or CPU, for the detector's onnxruntime backend. CUDA unless the detector is erroring.
• face_analysis_det_size - AUTO, which tries 640, then 320, then 160 until a face is found. If small or far-away faces aren't being detected, this is the knob.

Outputs

The three outputs - FACE_DETECTOR, ARCFACE_MODEL, IMAGE_PROJ_MODEL - are all typed MODEL, but don't treat them as interchangeable. All three feed Extract ID Embedding; the face detector alone also feeds Extract Face Pose Image. Wire all three or nothing downstream works.

Installing

The pack is on the official Comfy Registry, so ComfyUI Manager is the easy path - search ComfyUI_InfiniteYou. Or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/bytedance/ComfyUI_InfiniteYou.git
cd ComfyUI_InfiniteYou
pip install -r requirements.txt

Restart ComfyUI after. Then you still need FLUX itself - the diffusion model, VAE, and text encoders, per ComfyUI's standard Flux examples. The README's example workflow assumes the FP8 versions of those, because it's tuned to run in roughly 24GB of VRAM (the bf16 route peaks around 43GB).

Where people get burned

• insightface/onnxruntime refusing to install is the classic scar of this whole ArcFace identity family. It's not you - every tool in this lineage has the same install pain, and a recent-ish Python/venv usually fixes what won't.
• The licensing asterisk. The code is Apache 2.0, but the InfiniteYou model is CC BY-NC 4.0 and the InsightFace weights are non-commercial research-only. This is a research stack, full stop. Don't build a product on it.
• You can't skip the downloads. If a workflow errors with a missing-model message, the first run hasn't finished fetching the multi-GB stack. Let it run once before you go hunting for the real bug.