FilthyDegenerate

The Simpsons Style [LoRA]

Added 2024-12-16 18:09:54 +0000 UTC

Third LoRA, faced a lot of issues with this one, seens they are undercooking, thus increasing the LoRA's weight is necessary, this is not a bad thing as burnt render is worse, but still I want to be able to trigger it with the default value aka 1.0

I believe it is the batch size that it is screwing me over, I want a higher value to save time, but that saving perhaps is making the LoRA weight have less impact as If I used a lower batch size (1).

But I learned even more today, specially about how predictable is AI when it comes to training, if one is to pay attention to the samples one can notice that the first 10 epochs will give the 'rough' estimate of your character/style, is not a usable LoRA at this point, then at 18~20 epochs one will notice that second change, this is a LoRA that can be used, not as a final result, but it is perfect useable, at this epoch the AI should more or less learned the most common features of the character/style, after that the next 20 epochs are to 'refine' said details, thus meaning that on average a LoRA can be trained with 40 epochs, so for a fully working LoRA we are looking at 40~ minutes, for the first 'rough' we are looking at ten minutes, for a proper 'preview' we are looking at 20 minutes of training, this pattern is the same for anything one is training, be it character/style or concept.

the sweet spot for the datased is between 30 to 50 images, the epoch can range from 25 to 50, quality matters a lot, after 40 epochs the AI will be able to mimic the 'quality' of the image, so having a good quality dataset is far more important than a bigger one.

I found that concepts are far, far easier to train than characters.

https://civitai.com/user/BootyAttackArt