@Dr. Furkan GözükaraI have noticed that the quality of training on kohya is worse recently, is it th
@Dr. Furkan GözükaraI have noticed that the quality of training on kohya is worse recently, is it the same for you?
Fi am using older commit :d
Fi also updated readme file
Fshared in my video description
Pthe latest commit works for me
Fworks but dont learn
Fno errors :d
Pit learns!
Fweird didnt learn for me
Fi even opened a topic
Fbut i did a fresh install
Fdidnt use older venv
Pdid you use torch 2 or 1?
Fi think torch 1
Ptorch 2 worked for me, using rtx 3090
Fi should retest
Pif you could, try to use like a one week old commit and compare it to the most recent commit
Fi used my video commit
Palso this is not just local machine but even runpod also worse training quality
Yok i'm deleting it and i figure how to upgrade khoya and create a dreambooth with it but it create only one model and i seek a way to doing like automatic 1111 : multiple models so i can fiond wich one is not too much trained
AInteresting LR for Adan Dadaptation Optimizer
Fyou use buckets? or train for that resolution?
AVRAM usage is a bit hectic
Fthat's a lot lol ,worth it?
AProbably not, I am still training, it/sec is almost 1/3 of using Lion, so very slow
Nbut if they are different concept, then it doesnt really matter if you did 2 concepts at once or just one training session after another right
Fno i use single bucket. all images are 768x1024
Fit matters since you are also training unet
Fthe more you train unet = bad
NI see so in multiple concepts, it doesnt train the unet so theres some advantage
Fso you train that resolution? SD 1.5 model?
Frealistic vision yes
Fcan you share some parameters like LR, steps, etc? so i give it a try too
Fsame as in my video
Fhow many training pictures, how many reg images, that kind of stuff
F1e-7 lion
Fyou like lion
Ftried 90, 85 class images
F30 training images
Fthat's a lot of images
Fand how many vram does this need?
Fit is very slow
Fyou need 24 gb card
Fbecause i dont use xformers
For gradient checkpointing
Fso the a5000
Fis what i normally use anyway
Fno any 24 gb can work
Flike rtx 3090
Fbut a5000 is cheapest i think