Advice for training style with dreambooth using a large dataset

[TentacleSama4254]

I have a dataset of 11,000 images and I wish to train a style using these images. I'm wondering what settings and number of regulization images I should use for this

For now I was thinking of using

• 45 steps per image
• learning rate of 8e-7
• 25,000 regulization images
• constant learning rate with a starting factor of 0.9

Any advice on this would be greatly appreciated

yes, I know the training would take forever with this much data but I'm not sure if decreasing the trading steps would affect my training quality

[alelordelo]

Hi, Im also curious about DB on large datasets, but didn't find any info around...

Did you ran the training?