The sheet is the whole job: twenty or more stills of one face, varied angles and expressions, one full-height frame, even lighting, no sunglasses, no crop through the jaw. A weak sheet produces an identity that drifts, and no amount of downstream prompting recovers it.
Once trained, the identity becomes an asset the client owns. That is the difference between hiring a face and building one: the second one is still there next quarter, at the same cost per variant.
How many photos do you need to train an AI character?
Twenty or more, at varied angles and expressions, evenly lit, including at least one full-height frame. Variety matters more than volume.