This girl doesn't exist, but is created by prompting age, gender and nationality, creating 107 pictures, 39 of which are facial closeups from different angles using different facial expressions.
In other words, use this in good conscience if you just want a uniform, flexible uniform face without it looking like anyone specific.
Description
Different training, exact same dataset. Trained on a flat learning rate of 0.004 in 1500 steps
FAQ
Comments (5)
On some of them (especially in the anime models) the character looks european. Have you messed with negating some of the racial characteristics that might come up in popular 'mixed' model pictures?
I'm puzzled by that myself, really, because they're trained on images with the same face on every single one.
I do have a thoery, and it really is mainly speculation, that something was remaining in memory after training the Spanish generic, as I hadn't restarted the webui in between.
And between you, me, and everyone else who reads this, I occasionally go an image of a bearded man, and I absolutely guarantee there were no such images in the training set.
Uploaded version 2 with new training parameters.
@tvange365 - That would be hilarious if 99 images were the female, and you snuck a bearded man in there without effect... X) I think I've clarified that some of the reamining 'ghost in the tensors' happens when using the same seed value (in-spite of changes to the prompt). Each iteration biases especially greater away from the Neg that is used. I have yet to scientifically asses this (only my observational intuition). Maybe someone has written in-depth article pertaining to this?...
@tvange365 - I'll check it out.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.