Text encoding percentage

[jysung]

So a few days ago, was text encoding happening at 100% in the fast-dreambooth colab?

1. What exactly is text encoding? Any good documentation about what that is?
2. If I am training on a person with only 10 images, do you have any recommendations on what it should be set to? Thanks!

[TheLastBen]

Yes, before the update, the text_encoder was set to 100%

In simple terms, if it is set to 100% of the total training steps, you will get accurate representation at low steps count, but it will be harder to apply style to the trained subject, you will notice exaggerated overfitting.

If you want only photoreal images, you can set it to 100% and keep the steps lower (1500 steps for 10 input images).

If you want to stylize the trained subject, keep it around 40% and set the training steps count to 2500 for 10 input images.

[SAC020]

I am not clear either about what text encoding is. Does it require the input images to be named descriptively (such as "a woman holding a flower" or whatever description)? Where does it take the text from, how do you specify it?

[TheLastBen]

Yes, but you might be able to retrain in a limited way, depending on the prompt used

[SAC020]

Thank you

[TheLastBen]

happy to help

[SAC020]

Is it possible to have more than one subject in a model? If so, how should I name the files, and the model?

[TheLastBen]

Yes you can, you need to use a different name (identifier) for the images of the second subject and retrain the model