Face Analysis Models

This is the node you drop first and forget about - every other node in ComfyUI_FaceAnalysis plugs into its ANALYSIS_MODELS output, so nothing downstream runs until this one loads correctly. It doesn't do anything visible itself. It just picks which face-analysis engine the rest of the pack talks to, and that choice is 80% of whatever headache you're about to have installing this thing.

What it actually is

Face Analysis for ComfyUI is a small utility pack by cubiq - the same author behind ComfyUI_IPAdapter_plus, so if you've fought with IPAdapter FaceID before, you already know the shape of this dependency. Where IPAdapter FaceID generates a face, this pack measures one: crop it, mask it, warp it, compare it to another face and get a number back. The README is upfront about the flagship use case - send three reference photos of a real person through this pack, average them into a baseline, then score your generations against that baseline instead of eyeballing "does this look like them." That's what FaceAnalysisModels sets up for the rest of the pack.

One thing worth knowing going in: as of April 2025 the author put the repo in "maintenance only" mode. It still works fine - nothing here is broken - but don't expect new features or fast bug turnaround.

The two inputs, and this is really the whole decision

• library - insightface, auraface, or dlib. This is the actual face-recognition backend.
• provider - CPU, CUDA, DirectML, OpenVINO, ROCM, or CoreML. Just pick whatever matches your GPU (CUDA for Nvidia, CoreML on Mac, CPU if you have nothing else).

The library choice is where the real tradeoffs live:

insightface is the accuracy default and what most identity tools (InstantID, PuLID, FaceID) use under the hood - it's the ArcFace embedding, the same one that shows up everywhere in this corner of the ecosystem. The catch, and it's a real one: the pretrained models are non-commercial-only even though the library code is MIT, so if you're building anything you plan to sell, this path has a licensing wall baked in. Install used to be notorious too - years of no PyPI wheels meant a C++ build with Visual Studio tools on Windows. Version 1.0 (May 2026) finally dropped that requirement, but a lot of install guides floating around still describe the old painful process, so if you're stuck, check you're actually on 1.0+.

auraface is fal's Apache-2.0 answer to that exact licensing wall - commercially clean, competitive with InsightFace, worse on nothing that matters for most workflows. The catch here is the model isn't auto-downloaded: grab the files from Hugging Face yourself and drop them under models/insightface/models/auraface/, matching the folder InsightFace itself expects. Miss that step and FaceAnalysisModels will fail to load with a missing-model error that has nothing to do with your ComfyUI setup.

dlib is the oldest and lightest option, CPU-friendly, and doesn't touch InsightFace's license problem at all - but it needs four separate model files (two shape predictors at 68 and 5 landmarks, an 81-landmark variant, and a face recognition ResNet) manually placed in a dlib folder. It's the one to reach for on a machine with no real GPU or when you specifically need the 81-point model - the two other backends don't offer that landmark count, and a couple of downstream nodes in this pack (FaceWarp in particular) can use those extra forehead points if they're available.

Installing it

• ComfyUI Manager - search "Face Analysis for ComfyUI", install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/cubiq/ComfyUI_FaceAnalysis, restart.

Manager gets you the node pack itself, but the model files for whichever library you pick are still on you - none of the three backends are one-click.

Common issues

Most failures trace back to a mismatch between what you selected in library and what's actually sitting on disk. Picked dlib but only downloaded the 68-landmark predictor? Fine until a node asks for 81 points. Picked auraface and skipped the manual Hugging Face download? Load error at startup, every downstream node greyed out. If you're on insightface and hitting a build failure, check your version first - 1.0+ shouldn't need a C++ toolchain at all, and if you're still being told to install Visual Studio, you're likely on a stale install guide, not a broken environment. And since the repo is maintenance-only, don't expect a GitHub issue to get you unstuck fast - cross-check against the IPAdapter_plus install docs (same author, same dependencies, much bigger user base) before filing anything.