[EGreg]

Actually, there are many ways. For example they change colors on your screen and check in real time how it reflects on your face, eyes, etc. Very hard for a model to be trained to respond this quickly to what's on the screen.

They also have you move your head in multiple directions.

[cocoto]

You could always generate a random face model with real time rendering with enough details to trick any AI detector (or even human) and then you can do real time animation to orders or screen light tricks. You could also simply use some face filter on your face and these ones are really convincing these days (like on Snapchat and such).

[EGreg]

Show me such a model.

It would be interesting to see a model completely indistinguishable from a real human in behavior, as well as real-time reflection off different surfaces, etc.

The next step would be to make a complete digital clone of a person based on surreptitiously recording them with hidden cameras. I doubt it's possible.

[viraptor]

The pieces are there. If you're not modifying everything in the image all the time, there's no reason to run it through a visual model. Generate it once (we have it), transform into textured 3d model (we have it), animate and map to movements with vtuber software (we have it). Adding screen colour reflection is trivial. We just need a pipeline for this.

We had facerig for over a decade now. Facefilter recently. It's not hard anymore.

[ddtaylor]

This is doable using high end stuff like Runway with a draft quality.

Your better bet would be to generate a face as an image and then you can easily generate that same face in different expected poses and conditions. You can then use existing models where you get to select the starting image and the ending image. Add some filters and noise to just make it look like normal crappy low light camera.

As for the color that's another expected condition and can be overlayed or pre-generated.