Best way to make flux learn new people

[michaelcyshield]

hello, I am trying to fine tune flux to be able to generate pictures for articles about celebrities it currently does not know how to draw, so for example if this celebrity got a new reward I want to be able to generate an image for that. How can I go about this in terms of the data?
currently I am trying to get images of these celebrities where they are the only ones in the image, is this necessary or can It have multiple people in the image

secondly, I am captioning using blip3 and instructing it to keep its captions quite short as per the advice I read in forums for fine tuning SDXL so for example it would generate outputs like Megan Fox giving a speech

I used to have in my dataset these people doing things with other people like shaking hands and stuff but it seems to not be the best way to go about it.
I also hard code the name of the celebrity at the beginning of the prompt because blip3 sometimes omits the name from the resulting caption desplite how much I emphasize that it needs to mention it.

does anoyone know how I can go about this task as well as any good learning rates for this and other training configs such as the config for the lycoris and such? I am really hitting a stump here