The best way to evaluate generated faces is to first send a batch of 3 reference images to the node and compare them to a forth reference (all actual pictures of the person). That will give you a baseline number that you can use to compare to generated images.

Installation

Please download the DLIB Shape Predictor and the Face Recognition models and place them into the dlib directory.

Precompiled Dlib for windows can be found here.

In this repository you also find a workflow that uses IPAdapter to generate a few images and return the distance to the reference face.

The extension of course requires dlib that has to be installed into the ComfyUI environment.

Important notes

There are many ways to do this. At the moment I'm using DLib as it's fast and easy to use, if there's an actual interest I will release more options (insightface?).

Also, I'm not an engineer and I don't know what I'm doing, hopefully someone more experienced can chime in.