There's IPadapter for SD models which only requires a single photo at runtime, though I don't know exactly how strong its adherence is. There's also training a LORA which would only require a few images and take maybe a day at the most, which would be an easy investment of time and resources for someone who wants to churn out tons of images with that one face on them. And of course there's the simplest method of all: just crudely pasting the desired face into the image and then sending it for a low denoise img2img pass to make the inscrutable machine integrate it into the image better.

That level of extremely basic and rudimentary technical knowledge is probably beyond the influencers themselves, but certainly not beyond whatever flunky they pay to maintain their brand like that, at which point it's probably self-preservation and not wanting to create an automated process that can replace themselves that stops them from using that approach.