If instead you have a cohort on hand — -i.e., you do not want to send your data to a third party for any reason, or perhaps you have energetic undergrads — -then you could alternatively consider local, open-source annotation such as CVAT and Label Studio. Finally, nowadays, you might instead work with Large Multimodal Models to have them annotate your data; more on this awkward angle later. - Source: dev.to / about 2 months ago

I’ve done something similar, using AI Generated Photos run through ImageMagick to reduce and pixellate the images, making them suitable for low-res/retro games which demand a wide variety of “people” portraits. Source: about 2 years ago