Training is suddenly much faster.

[hbqdev]

Did anyone notice the same thing?
I'm on the 3090.

[1025307336]

*2,But the problem is that the training effect is very poor

[ExponentialML]

Edit This is now merged, so you can just update the extension.

Hey @1025307336 .
Could you test the changes here in this pull request? 001836d
Simply change the .yaml file after converting the checkpoint models/dreambooth/*your_model_folder*

1. Find the .yaml file (will be the same name as the model folder.yaml).
2. Make the changes like the link above to your_model.yaml file (red = old, blue = new).
3. Restart the entire webui to make sure the model is reloaded in the Dreambooth tab correctly.

If it does work, then you can change all three of those .yaml files to match in your extensions/config/*.yaml directory to match for the time being. That way they're automatically created when you have a new checkpoint.

Report back and let us know if the results are better, thanks!

[rpgsolomaker]

Did anyone notice the same thing? I'm on the 3090.

Would you mind sharing the settings that you used? And how many images were in your dataset directory and classification folder? Ty

[1025307336]

I've found out why，His counting algorithm in cmd has changed，steps=bs * Normal running speed，The bs setting may have been disabled so it didn't slow down the computation