Avatar
:VD15_0::VD15_1::VD15_2::VD15_3::VD15_4::VD15_5::VD15_6::VD15_7:
e19581db068951fff6de5ca917ab89a72b2aa323a063927e9a6ad9500eb5617d

Later models were very crispy, but some of the earlier checkpoints were pretty nice. At least I know I don’t have to train it for multiple hours to get a good result. The sweetspot seems to be around 1000-3000 steps.

I’ll adjust the dataset some more and test again later today

It’s not terrible but it’s not particularly good either. It’s definitely picked up on how the artist does hair and shading. Faces are roughly the same style, but a lot of the output is rather mishapen. (first attachment is from the training set, rest is output from the LoRA)

IDK what to do to make it better. Not sure if I should go back and write better prompts for the training data, add more images to the training set or just run it for more epochs

wtf happened to lightworks? Did they get bought out? They took their source code private and are only publishing builds for w*ndows. The versions that are still 'Free' require you to create an account

Neighbors are lucky I’m not Chinese or their dog would be in a stir fry right now.