A LORA model trained on Jake Orion.
To use place the words "jakeorion person" in the prompt.
Description
Initial release.
FAQ
Comments (3)
Good evening, I would like to know how you manage to train with more than 24,000 steps, your models don't fry on power 1, do you reduce the training rate so as not to "fry" the photos and how many epochs would you recommend? thank you very much if you answer.
Hi. I tend try to use +200 images at least and very low learning rate. For example with this LORA I used 438 images and learning rate of 0.000684. My formula is 0.03 / # images. always in 512x512 resolution as is the native SD 1.5 resolution. To get the images ready I start with trying to get high resolution pictures (Example: 1800x1400) then uncrop to get a squared image (1024x1024 from uncrop) as normally it gets a full body picture, the I crop the rest of the picture: 1400x1400 then 1024x1024 then 768x768 then last 512x512 so from 1 high resolution images I end up with 5 images all scaled to 512x512. Look at it as a digital zoom in. This makes the LORA more flexible and less likely to burn.
I use the automatic1111 dream booth extension as I work on M2 Mac so other methods doesn't works because they require Nvidia and windows.
@eddnor Hm, thanks for the tips, I work with few real images, but with 30 steps of 10 images which is approximately 400 per block, but I notice that they burn more on the body part, it would be a good tip to use the "body" in the creations, but I will try to reduce the learning rate and use 512x512 to reach 20 thousand steps!! Thanks for the tip!