Thanks! With flux, I use close-ups when I can, but in cases where the face play a role, like this one, I hide some face features with random colored geometric forms(green square, red rectangle, blue circles, etc) the caption system inderstand I try to anonymize the persons and the geometric features is in the caption, like eyes censored by a red circle. I try hide the traits I don’t want in the face while keeping it as random as possible(I don’t hide only eyes, but a few mouths or nose here and there). So far I’ve seen flux replicate one of these once in epoch samples, but it seems to work very well to keep it from replicating likeness. There’s a few things I do to keep a nice flexibility which is well documented in various how-tos; each epoch is tested with the same threee prompts which aren’t trying to replicate the subject in dataset. If the subject is mostly thin in dataset, I put an obese in the samples to see if it starts to overfit. Also, I try to keep my dataset with various subjects style which are doing the same thing (the thing I want to train), not too many angles, not too many variations, make sure it’s clear what’s happening and the thing you train on is the first thing that comes in mind when looking each images.

@LatentDream Thanks! For the dataset, I don’t own every images I use, so I can’t always share it. I don’t mind sharing or writing an article, it just takes time to do it properly, which is time I don’t have to generate. I’m far from an expert, I still experiment and learn every day, so I feel I don’t have a solid base enough to do a whole « how-to », but the way I hide features, I feel it’s working wonders and maybe I should do an article about it.