parameters look fine, however you may have to play around what base model you are using
parameters look fine, however you may have to play around what base model you are using
Hyes i was trying if upscaling my training images and Changing base model would improve the end result just now
Fi just found something new testing speed difference
Fto runpod web ui
V@Furkan Gözükara SECourses It would be so good if you prepare a tutorial about how we can make a clone POD through container/Docker. It's so useful for everyone
Fthe only clone way is using runpodctl
Fso you can send your workspace folder entirely
Fwhich i used
Fbut i can make a short video
Fusing zip command is better
VI couldn't understand how we can compress the entire SD folder through runpodctl
Fit does but slow. so using zip command better. for that i can make a video good idea
Fit also requires same equivalent of free space ofc
Vright! so for compressing and decompressing what command we should use ?
Fi dont remember from memory i have to look
Fit also requires some extra script install
VThx, another suggestion. if youtube give you an option to change the font color for your video caption please change it. with the current font style it's hard to read
Fyou can do that. that is available for everyone
Fclick options
F
F
Danyone ever have auto1111 dreambooth extension throw errors on individual files after cacheing latents , i got to 1 percent of training and threw error about an indibidual file 1400 image datset, glad i caught it before sleep ....... anyways wasted hours of time and just killed instance , kept throwing errors on seemingly random imgs in the set
Fi seen that
Fimage could be corrupted
Dstable tuner works so much better about just running , but the current scriipt stable tuner exports is currently broken for runpod
Fout of resolution than others
Fimage description can be problem as well
Dhmmmmm
Dits possible uploading was funky
Dsince my internet died halfway and i went back and did ignore replace on ones that were already there
Dhowever when i viewed the files in question on workspace they appeared fine
Di think im going to stick to 200 image fine tunes and rely on merging
Dtoo annoying with larger dataset
Danother strange thing is i have been able to train 1.5 on 768 resolution with batch size 2 before , while last night it threw cuda oom
Don 3090
Ddr F whats your preffered card for training? i was using a40 on vast but datacenter in sweden doubled the price
Frunpod has great cards
Ffor higher resolution
Fwe did 1k
Falso use runpodctl send command
Fto upload
VThx but I'm talking about the font style for video caption, not subtitle
But I found alternative way by scrolling through the right window
But I found alternative way by scrolling through the right window
Fah i dont know if that is on my hand :/
Dis there any use in multi gpu instances for training?
Dafaik theres no way to use multi gpu
Fyou cant do that right now with web ui
Fperhaps with diffusers
Fbut much hard to do
Danyone mess with visual chat gpt yet?
DGitHub
VisualChatGPT. Contribute to microsoft/visual-chatgpt development by creating an account on GitHub.
Dis it possible to set up visual chat gpt to run on a phone number you could text / mms ?