why is torch 2 slower on kohya ss?
why is torch 2 slower on kohya ss?
FYe that is interesting. Maybe because I tested with Fp16 and not bf16
FCan you also test?
Ii selected no there
Ii have 24GB VRAM
FYes makes sense
FMaybe that is reason
Ialmost ready to train, working through your tutorial, thank you
Ihow are you running stable diffusion and kohya ss at the same time, it told me i could only run one at a time
Invm got it
IHey guys, i need to use SD+ControlNet+ Temporal Kit, do you have any colab i can use please? i used colab only for dreambooth
FChange port 
FUnfortunately don't have colab for that :/
IThanks anyways, i'm doing some test for making a decent video but i need more vram
Imy 8gb are decent but with more i can have more stability
Ilol now to gen 600 images of "man"
IIthank gods for 3090
DI'm going to try InvokeAI and see how that works on my machine. So far I like it's installer. We'll see how it goes...
JI don't find the same at all, without captions, it's very difficult to try and align the training, seems to wonder all over the place
CHey @Furkan Gözükara SECourses, are you open to consulting inquiries?
CRunning a commercial stable diffusion inference pipeline and we need our DB training audited.
JI used unipc and cranked out a 100 batches of 5 at like 10 steps
JIt worked pretty well
JTook about 15 mins
Ii just did default and it genned 600 in about 5 mins
I20 steps, euler a 512x512, not bad
Iim only to the settings part now lol
Iif i train a single lora before bed ill be happy
DWhy not just download the images of Man that Dr. G. made available for download as a Zip to save time, or are you testing your system?
Iam i doing something wrong? its saying its going to take 1.5 hrs, i thought the loras only took like 10-15 minutes?
Iit said it was behind a patron wall
F8 gb good for lora
FYou are teaching a person? How many images you are training
FYes I am
JYeah, myself, I've tried between 10-25 - Lora has been a complete failure, TI gives better results, but still not perfect
JIt's strange, 6 months ago I trained some really successful TI's, but now I can't seem to get a decent result
FCan you try my latest video?
I1.27 it/s isnt that slow?
FI have to sleep now hopefully see you tomorrow
JHave done, not captioning my data set seems to be the problem, will retry tomorrow
FKohya is slow
JTake it easy mate
Iso it makes sense for it to take 1.5 hrs to train a lora on a 3090?
FI find that sd web ui faster
FMy trainings were 45 min
FBut it depends on number of total steps
FAnd batch size
Iwow, TIs are much faster
FTi really fast
FI am afk now
