IllusPony

This is a merge of Illustrious 2.0 and Pony Model, Aimed to remove the cons of Illustrious 2.0.

Hot take

Honestly, Illustrious 2.0 is overhyped, it's only trained on Danbooru Image Dataset, which makes it so that it's highly viable for making Pornos, however its only trained till Epoch 2, on 20M dataset with batch of 512, making it so that it doesn't excel at any single thing, its jack of all trades, can't even pose properly. its highly biased towards artist names so you can use it. Since, there are tons of old images which makes it so that, the actual rendering comes out to be cell shaded color which is bad for anime model. With Pony, it overcame it a bit however, it needs to be trained properly, it's name doesnt actually means what model tells. During the Cooking the merges, I tried with different layers, lots of model composition were tested, and Honestly, Half of the layers were useless as they were generating NaNs. Only a few layers,

MID 1,2,3 in XL, these 3 layers are responsible for Style, Shape, Color a bit

Out 7,6,5,4 In 9,8,7,6 You can work around with these,

Overall, making it so that most of Loras(and varients) trained on Pony, Autism, NAI etc. can't do shit.

It's also because Artist name is contributing a huge part in training the model.

They decided to use 4e-5, for learning, its a cool number mid-ranging between training for Character and Style, however it only learnt the styles of artist, so there's this Ethics issue, as you have to use "Art style of {XXX}" . Making it so that Umbrella terms are unusable.