Thanks anyways, i'm doing some test for making a decent video but i need more vram
Thanks 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
CI just added you. Will give you more information about the project there
Iits much faster when i switched to fp16, about the 45 mins you said
I3.6 it/s instead of 1.24
Iand our speeds should be similar you have the TI version of my card
Iits also only using 30-50% of GPU isntead of 100% while going twice as fast, nice
Inow my room doesnt have to become a sauna to train a lora, lovely
Iresults arent good, even after 10k steps. i think its my input pics
Cis dreambooth the primary way that people train loras? or is there a different way to do it if you have a larger dataset?
Vhey guys, after the recent google colab limitations for stable diffusion webui, does anyone have a workaround for it?
FInput pics and the prompt matters a lot. If you do raw prompt you won't get good results
FLora training is based on dreambooth training. It is an alternative way
FYou can do larger training. For example if you don't use classification images and use file captions it becomes fine tuning
FYou can use no web ui scripts have you tried?