sd 21 is really hard to train and get good results
sd 21 is really hard to train and get good results
Finteresting i think i should compare
Fby the way it turns out that 768x1024 loses styling ability significantly
Fgood for real images but not for style
Fand my training dataset had so many repeating cloth and background
Fthat degraded styling significantly
Fawesome ty so much
T@Dr. Furkan Gözükara in cloning, can I resume training Eg i left at 2550.gpt can I start again from that?
Twhy is this picking random voices? Can i use one specific tone only? Please guide
Fyes you can continue training from there
Frandom voices could be related to your training progress and dataset
TDoes more training means better result? which parameter to select to resume training?
TDataset contains only my voice of 16 mins and trained it for 5 hours until 2155_gpt.pth
Fnot at all
F16 min voice could be low amount
Fusually more is better in training dataset as long as high quality
Awand.ai helps to determine where styling gets lost as well 
Fthe model learned even the whites in my beard haha
F
AThats awesome
Fbut there is too much memorization because of my training set
Fi should get good different bg and clothing pictures
AOh I thought you wanted it like that
Fit is good but cant stylze
TWhats the minimum amount of dataset? Also, will it resume if I start it like in video?
TI mean will dataset training resume?
Zthey need to hurry up aand make it so us lower vram guys can dreambooth
Zor should i just rent a runpod with someone with a beast card, train my models, then use them in auto
AYou don't need an 80GB beast though, 24GB is fine unless you want to use Adan D-Adapatation Optimizer, then you need at least 32GB
AW&B can also send email or Slack notifications when your training completes 
Fi trained with 2 hours speech
Freally this is 32 gb?
Fwow
AOnly Adan D-Adaptation Optimizer, not the other D-Dadaptation optimizers, the others can all run on 24GB VRAM
AMaybe its a bug in Dreambooth training for the DB extension, because someone is using the DB extension to train LoRA with Adan D-Adapt with just 8GB: https://civitai.com/articles/52/d8ahazard-dreambooth-extension-guide
RSo do you advice better to train with 512 x 512 pixel image and SD 1.5 model ?
Fsd 1.5 much easier better
Falso using public models are even better
AW&B also helps you compare 4 different trainings to determine the optimal average epoch so you don't overtrain
Faccording to your aim
AWell can be much more than 4, just I have only done 4 so far
FW&B?
AWeights & Balances (wandb.ai)
Fi see paid service
AYou can enter your API key into Kohya_ss and EveryDream2trainer, not yet supported by Dreambooth extension but I logged a feature request to add it
AFree for personal and academic use.
Rokay so can I use your 512 x 512 Reg class man images of your Patreon for training ?
AI am not paying, I am using it for free
Fyes. make sure that you are using the correct model
Fnice
