SDXL Lora training on runpod

[meanin2]

Hi, first off shout out to @TheLastBen I love your work!
Thank you so much for making training accessible for everyone!

A few questions about sdxl lora training in runpod:

For image size can I use different sized images and set crop to false? (Above 1024x1024 but all different sizes and ratios)

Does image count matter? (Or is it the same 10 you always recommended for the collab notebook?)

Where can I upload my created lora to hugging face? (I dont see a save or upload cell anywhere in the sdxl training notebook)

Is there minimum requirement for training a sdxl lora? (ram, vram, etc)

Thank you so much

[TheLastBen]

Hi and thanks,
yes you can use any size you want, make sure it's 1:1

10-20 images are enough to inject the concept into the model.

since LoRA files are not that large, I removed the hf upload cell.

3090 is good for LoRA training

[meanin2]

thanks, last question,
In the loras I trained so far, the file size ended up being around 1.3gb shouldn't they be smaller?

[TheLastBen]

reduce the dim to 64 or 32, but the quality will decrease a bit

[THX1139b]

So the default epoch values are unchanged for either 10 or 20 images? Or should I double both numbers for 20 images (240 + 80 text) ?Thanks!

[TheLastBen]

if the number of images is higher, reduce the number of epochs to keep the total number of steps in the 1500-3000 range

[THX1139b]

Aha, I have clearly thought about epochs the wrong way. Thank you!