do loras get better the more they are used?
do loras get better the more they are used?
VHello
VI'm pretty much new to this self-hosted AI generation, I've been working for the last week with Fooocus (using SDXL) and I manage to build a create some consistent images of my AI Influencer.
Fflux lora is amazing
VThen I came across Flux and I kinda fell for it, but I've been having a hardtime doing the swapfaces feature like I used to do on Fooocus(that automatically generates based on the prompt and a set of faces you provide as reference).
Fcompared to sdxl
VNow I`m using SwarmUI, but I have no clue on hoe to achieve the faces I have already with future generations. Does anyone have a suggestion of how can I go about this?
Fwhat you want to do?
Fonce you do training you dont need swapface
VYou mean make a LORA?
Fyep
VI assume I don't have enough computer power to do that
VHow many images do I need of my subject?
F15 good
Fyou can use cloud
FYouTubeSECourses
Ultimate Kohya GUI FLUX LoRA training tutorial. This tutorial is product of non-stop 9 days research and training. I have trained over 73 FLUX LoRA models and analyzed all to prepare this tutorial video. The research still going on and hopefully the results will be significantly improved and latest configs and findings will be shared. Please wat...
FYouTubeSECourses
Unlock the power of FLUX LoRA training, even if you're short on GPUs or looking to boost speed and scale! This comprehensive guide takes you from novice to expert, showing you how to use Kohya GUI for creating top-notch FLUX LoRAs in the cloud. We'll cover everything: maximizing quality, optimizing speed, and finding the best deals. With our exc...
Flower LR 256 images failed , it is undertrained at 40 epochs 
Vwhat those terms mean?
VSeem them before
FLR means learning rate
Fyou should watch above videos fully
F
Fthe resemblance not there yet
Falso lacking that lora brought realism
Fi will post 256 images result on patreon now
Fwill do another training
A
Ahands still major issue in fully body poses
Fstarted new training :d
V
Valso training here
Vbut on the cloud, my 3070TI has a shitty amount of VRAM for that task
Fuse massed compute
Fway cheaper
Fway better quality with our config
VI had some credits there
Fh100 total waste of money :d
Ffor example lora rank 16 = much lower quality compared to 128
Fi compared
VI'm training again on 128
Awhat quality specifically dr?
Vdid it and now all I get are big headed photos 
D@Lecturer - Main Admin I wonder if you've ever used modal.com? How difficult is it to run kohya_ss there?
XXSomething that I see has helped in prompts that include multiple people is changing what the people are wearing their age and hair color
MI'm still waiting with Flux for several reasons. For one thing, what I see from current loras is that low weight/alpha is enough for posing, so some people use 8s. On the other hand, human anatomy is horrible, so not trained on nsfw images, so it will be a lot more work to train all body parts, and I'd rather wait for that until more sophisticated flux models are available. As I see it, it will be at least 2-3 months. The Pony is also being made for Auraflow first (which is also because of the license), so the transition to flux is a long way off for me.
WDoes fp8 vs fp16 lora quality diffrerent much ?
Fquality of training
Fa little bit