Just a fast tutorial for my character/person LoRa creation in 1.5 by using khoya_ss
First of all dataset, good resolution images, the higher, the better. Doesnt really matter if they have different dimensions, just a high resolution is the thing that matters most.
Quantity and variety of the dataset, I always try to aim for 20 images, I will explain a bit later here why. Variety, I try to get a decent mix of headshots and upper body/full body shots. Dont add blurry or unfocused images, you can unblur images with certain programs/sites if you have no other choice. Try to add variety in poses and clothing, just to avoid overtraining certain aspects.
Captioning, I mostly use blip captioning of khoya_ss under utilities, prefix I used to type the name of the character/person, nowadays I just use something random with no known value (like ohwx), I dont tend to do anything more with the text files, but SOMETIMES I try to add extra descriptions if I dont want to associate something that appears on most of the pictures with the person (eg necklaces, piercings, tattoos, repeating backgrounds)
Settings, and also the reason why I aim for 20 images. I try to aim for 1500 steps in total (to begin with). This includes the repeat with training images ( you can find training images online ). My batch size is always 2 (again this just works the best for me) and I do repeat steps of 25 per image and 3 epochs. So this means (25*20*2*3) / 3 = 1500 steps total. So that's about 6-10 minutes to train a LoRa. If you have less images than 20, adjust your epoch or repeat steps so you still have around 1500 steps. I used to do more steps for a better match in appearance but the LoRa tends to be less flexible if you do that. I advise to just use adetailer on the face with a higher value instead. Personally I dont go for 100% accuracy, HOWEVER if this does not suffise and you want a closer appearance, I would advise to raise your epochs first, and make sure save every epoch option is enabled so you can compare which lora works the best for you, this way it is easier to see when you are overtraining your LoRa. Again the quality of your dataset is the most important factor.
So that's about it, I will add my khoya settings in the attachments as images, feel free to leave questions or remarks. Also if anyone has a good tutorial for XL training. Let me know :-).
