i do think Timson have the right answer, i had similar results with overtrained Loras, but in the en
i do think Timson have the right answer, i had similar results with overtrained Loras, but in the end, it is mostly the resolution.
Mtrained a LoRA for flux exactly like the tutorial (for human images), the results are nothing like the trained pictures, and its not even a consistent character, just keeps pulling out new character as if i am not using any lora
Mand even tho i used hundreds of images, the size of each of the lora files was only 18MB
Fstripes happens when you upscale with lora
Ffine tuning doesnt have that issue
Fnot our config 
Ddoes the sort by similiarity app run on the gpu or the cpu? mine is using 93% of my cpu but 3% of my gpu?
Fit uses gpu
Fbut you have to follow requirements
Fespecially cudnn libraries
�@Furkan Gözükara SECourses Will you try flux fill training? or can you make installers for training it?
Fit requires 2x 80 gb gpu
Fi dont plan it
Fonly for client projects if someone wants
Funified donwloader updated
FF
�Can you make installers though?
Fmaybe
Flol i loaded lora config into dreambooth tab inaccurately and it broken the config
NYes. 0.8 dinoise
NYeah, it's high resolution, okay, I'll keep that in mind.
What tools do you use to smooth it out? Upscaler or something?
What tools do you use to smooth it out? Upscaler or something?
F2xh100 FLUX LoRA
TI was not able to fix it, just lowered resolution and then upscale with supir
TIf you need extreme resolution you probably need some tiling technique, as someone here presented
A@Furkan Gözükara SECourses do you have supir upscalerdownloader for massed ?
Fsuper fast to isntall on massed compute
FPatreon
Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon
Fmax few min
Aamazing
Aany tutorial for control nets?
Fi have this very through https://youtu.be/3E5fhFQUVLo
YouTubeSECourses
I show how to install Automatic1111 Web UI & ControlNet extension installation from scratch in this video. Moreover I show how to make amazing QR codes and inpainting and out painting of ControlNet which are very similar to Photoshop generative fill and Midjourney zoom out. Furthermore, I explain and show what are Canny, Depth, Normal, OpenPose...
Ajust amazing
Ait's the same concept for flux too right ?
Fflux controlnets are weaker but yes
Aso what approach should I choose I want a specific pose for example that's hard to imitate with just generation?
Aalso @Furkan Gözükara SECourses is super realistic inpaint a thing in Flux?
.@Furkan Gözükara SECourses which dev branch of Kohya allows Flex1 training and do I have to change any parameters for it?
F1 : not sure how to do this flux
F2 : what you mean?
Fi replied. same as we trian flux
Fsd 3.flux1
.I've trained OpenFlux and re results are very good but it took much longer training than Flux dev dedistilled.
FRTX 5090 faster than L40S with FLUX at BF16
Fi am using kohya and it was even faster :d
Fbecause flex1 has lesser parameter
Falthough it probably need a better learning rate
Fwhich i dont plan to research atm
Fonly if a company or client requests