thank Dr Furkan i will try, still the use of 1024 images for training is the reocmend, or 512?
thank Dr Furkan i will try, still the use of 1024 images for training is the reocmend, or 512?
Fbest
Flower quality degrades quality. i am even gonna test 1280 and 1536
Fit says this is a
Fi am gonna remove it from captioner :d
FHello, I'm getting back into model training. I am registered on patreon, is there a tuto to use masse compute please?
FI want to try Kohya Flux Lora to use them in comfyui
Fuse it on comfyui
Saafter clik in Windows_Install_Step_1 which file i must select after that
S
FIt's ok I found it
Fyes it is 1 click on massed compute now
Fand welcome
Fi think it should work
Fbut i prefer 3.10.11
Fworking perfect
Si must click this also right ? Fix_For_FLUX_Step_2
Si coudl install python 3.10.11 and dont have erro with forge and another isntallation?
Fit works with all
Fye first do a fresh install kohya
Fand then step2
Fuse my installer it switches to accurate branch
Sis already install and open, do you have some guide dr Furkan? i been working with onetrainer has been pass some time since i use kohya
Fi have old one but new one will be very soon hopefully
FYouTubeSECourses
In this tutorial, you will learn how to install Automatic1111 Web UI for SDXL. How to use LoRAs with Automatic1111 SD Web UI. How to install Kohya SS GUI scripts to do Stable Diffusion training. How to train LoRAs on SDXL model with least amount of VRAM using settings. All of the details, tips and tricks of Kohya trainings. How to do x/y/z plot ...
DI'm testing kohya on WSL how I can update to the new cuda libraries? I have to do it manually? is training but very slow is consuming more ram than windows
Fwindows slow- same vram
Fwsl was too hard for regular users so i gave up looking to it
Dyes but is working for me, just I want to know how to update the cuda libraries on linux manually in kohya
Fi used cuda 11.8
Fworks perfect speed on massed compute
Fjust upgrade to torch 2.4
Fi am afk
Dahh so is torch and not cuda?
Dwhat it doubles de trainig speed?
Dthanks
G(train_data_dir must be the parent of folders with images) Error
The training guide does not say how to properly name the folder. Please advise
The training guide does not say how to properly name the folder. Please advise
Ni use the dataset preparation section... set instance prompt (token word like ohwx) and class prompt (like man). then set the training images directory to where your source pics are and set repeats to 1. then give a destination directory (where your training images will be copied in the format kohya needs, and where you want your output loras to be written). then click prepare training data. then click copy into to respective fields. that will copy your training images into a new folder for kohya to use, like "1_ohwx man" - then when you start training, kohya pulls the repeats, instance prompt and class prompt from the name of the directory.
Dgreat speed on wsl and rtx 3090
G@Dr. Furkan Gözükara Bro you should add these steps for noobs like us, I had no idea about this step
GTraining is started now
Edoes anyone know the best way to upscale an image while fixing strange artifacts without adding new ones? I've read about using an upscaler followed by img2img with flux.What's a good and efficient way of doing this without spending hours on inpainting?
GUpscale inside Comfy UI. I do 20 steps a 1.5 X, it does an amazing job imo, sharper, more details, fixes most artifacts and increases the resolution size 50%, I save as .JPG and the size is around 500KB instead of 1.5-2MB while maintaining quality
Fhello when I'm on kohya on massed compute I don't have this part
GThis is probably a LORA on img to img with a low denoise, essentialy the photos are stolen from real women and they swap the face bro
GGot it running and still get an error. I dont get it man. I did all the settings for 24gb and 48gb both still fail
Greinstall Kohya, you have a different version, i did that and that reappeared. Also, make sure youre on the LORA tab not Dreambooth
TIt is perfectly possible
GSDXL can do this
GFlux with a Lora and your set