2nd version: bigger dataset, better configs, it should just be better overall :).
Cheers!
Feel free to chat to me if you've anyone you wish to make a commission for, I'll do anyone if you can provide the photos :D
====
My first lora!
I hope you guys like it! The trigger word is GA1. You can play a little with the weight, 1.0 is fine, maybe for full body images consider 1.05 or so.
I want to credit the guys at the discord of OneTrainer because they really helped. If you wanna make your own loras, and you think Kohya is like a brick wall, try Onetrainer, it's actually pretty intuitive by comparison, though you'll still have to do some reading. https://github.com/Nerogar/OneTrainer
Over at their discord, they were super nice and helpful. Kudos to them!
Recommended resolution: 512x768, adetailer, hires fix for bigger, and it should really be flexible, it's trained on base SD1.5, so, it should work on most checkpoints if not all. I tested mostly with CFG 7-8, but I never really go below those ranges.
Limited testing shows it should work fine with other loras too.
Enjoy!
Description
+Increased dataset.
+Weights now float32 instead of 16.
Results: it should be more precise, just as flexible, and honestly give you better results.
Obviously it's not faceswap, so you'll still cherry pick a bit, but it's overall -better-!
If you're getting too many portraits, just prompt against it a bit, it's flexible, it works.
I made all images with adetailer and hi res fix, but adetailer alone can get it.
Enjoy!
FAQ
Comments (4)
Good job - that's a definite fix from my POV. Thanks.
Ty! Happy you like it! Do post some images if you make them :D
I've uploaded a new file because the latest iteration isn't even worth a new version, it's just a minuscule change in the tagging. I changed the tag word, feel free to download it if you want a 0.03% better lora =P it's a realy tiny improvement so feel free to skip tho.
(turns out adding "woman" at the end of your trigger word turns your trigger into saying "THAT woman" rather than "GA1, a woman", and it's a little bit smoother for it.)
Weird, version 1 worked fine for me with ADetailer, but version 2 comes up with weirdness that you would barely call a face.
Thats weird... maybe share the prompt, checkpoint, adetailer prompt and resolution? I was making images with her yesterday and ut works fine, even repeating the lora. 512x768, adetailer as usual, ans hires fix for bigger.
As a note the lora is trained on 768 resolution, so if you want bigger than that use hires