okay thanks! one of these options would increase training speed?
okay thanks! one of these options would increase training speed?
Fi would chose bf16
Fman been working whole day
Ffull real pictures classification dataset
Fcoming to the patreon
AYeah, I don't understand it either, because I am pretty sure you said you were using that GPU type in RunPod 
AThis is awesome, will it working better than the usual classification images?
F
Fi am hoping so
Ffrom very high resolution
Fi will post different aspect ratios
F512x512 , 768x1024 , 1024x768 , 640x960, 960x640
F512x768, 768x512
Fthis can be used to fine tune as well
Ssaw a reddit post somewhere by someone that swears by using real images for classification images, instead of SD generated images
SI wonder how much of a difference it makes
SI think they mentioned using unsplash images as well
SGitHub
Flickr-Faces-HQ Dataset (FFHQ). Contribute to NVlabs/ffhq-dataset development by creating an account on GitHub.
Fi also collected from unspash
Swould be interesting seeing the results
Kdo you know how to find 2400 Photo Of classification images for women ?
K@Lecturer - Main Admin
K@Dr. Furkan Gözükara
F@dbai64 sure
Fdo git checkout commit_hash
Fi dont know have that
Fwow, my article is already on the second gogle page for the search "how stable diffusion works"
FXD
Falmost on page 1
AThat's awesome, congratulations!
Fdmn this aspect ratio took a lot of time
Ffinally fixed the script
Faa.zip4.92MB
Fi had been struggling to figure out like 4 hours. constantly debugging haha :d dmn
ATemplate has been updated:
Updates:
Updates:
- Automatic1111 Web UI version 1.3.1.
- Dreambooth extension 1.0.14 (no longer dev branch).
- Kohya_ss added.
- Tensorboard added.
F5 instance processing class images
Fspent 2 days for this dataset :/
ABy the way, you don't need to kill the web ui process to view the logs in my new image, you can simply tail the logs at
/workspace/logscomputer crashed :/
JAwesome, I'm really looking forward to that new workflow! Thanks for replying
Fsuper slow training without xformers :/
F
Fi am going to sleep hopefully see you tomorrow
MDoes anyone know where to get this ui for Controlnet
M
MNevermind, i figured it out. needed to update CN
AWhat is the best commit to use for Kohya? the master branch seems different from the videos and can't get it to work
AI find it interesting to watch the learning rate of the Facebook D-Adaptation optimizers
M@Dr. Furkan Gözükara I tried a dreambooth training with 2 concepts
1 of a female character and 1 of aesthetic
However, I feel my aesthetic is not trained well as it does not really generate my trained female model in that aesthetic, it is a hit and miss, I am asking here as I was watching one of SE courses videos and that part of style training was skipped.
What kind of settings are appropriate for style training?
And should I train my style and character separately and then merge the checpoints?
1 of a female character and 1 of aesthetic
However, I feel my aesthetic is not trained well as it does not really generate my trained female model in that aesthetic, it is a hit and miss, I am asking here as I was watching one of SE courses videos and that part of style training was skipped.
What kind of settings are appropriate for style training?
And should I train my style and character separately and then merge the checpoints?
Fresults are amazing quality