haha

thanks. I believe there are creators that published loras for that. or is there something missing in your view?

@dsszeo I'm not out to compete with other creators, but I don't mind doing loras if there's an interest.

thanks!

@dluckye I've been postponing writing some tutorials to post but have been lazy.

most info out there on loras is not applicable for most use cases. myths like: you need to use 128 dim/alpha, you only need 30 images, you need to crop 512x512, etc. most only applies under certain circumstances. for most loras you don't need to go beyond 8,16 or 32 dim/alpha (i.e. at most ~30mb). certainly for anime more than 32 dim is an overkill.

I do not remove any info from the metadata about the training. Meaning that others can learn from my files. You simply need an extension like Additional Networks in auto1111 webui and use it so see the training metadata.

most people fixate around the parameters but the #1 thing you need is quality dataset: in terms of representing the concept, a good amount, and flexibility. not having that right won't make it work regardless of how many different parameters you throw at it.

for this one I used around 500 images (no cropping or resizing!) with the subject matter clearly perceptible even in a thumbnail. I trained on koyha ss gui with 10 repeats, 512 resolution, 10 epochs, 16 batch size, 0.0001 LR, Adam8bit optimizer, gradient accumulation 1, keep_tokens 2 (both triggers)

I could share the dataset, but honestly you won't learn from it.

@dluckye 1st autotag with WD14 and then bulk edit tags with BooruDatasetTagManager (tp add trigger, to prune unnecessary tags, etc.). how much I edit tags depends on the use case

hey there. the payoff for doing SDXL is still small IMO. maybe at a later date