what is the difference between train_dreambooth and train_dreambooth_lora?

[Podidiving]

Hi! 👋

I was looking into examples and got stuck with a question:

there are 2 scripts: train_dreambooth.py and train_dreambooth_lora.py
What is the difference between these 2?

As I understood, train_dreambooth_lora just trains small lora adapter instead of full unet checkpoint. But then I got puzzled: if that is the case than what is the difference between train_dreambooth_lora.py and train_text_to_image_lora.py ?

What am I missing? Can somebody help?

Thanks in advance ❤️

[asomoza]

you're almost correct in your assumption, train_dreambooth_lora trains a small dreambooth lora.

The difference between train_dreambooth_lora.py and train_text_to_image_lora.py is that the first one uses dreambooth and the second is just a finetune lora.

[Podidiving]

trains a small dreambooth lora.

this confuses me a bit.

• train_dreambooth_lora trains only lora adapter weights + unique identifier, but uses dreambooth training scheme (i.e. uses prior preserving loss).
• train_text_to_image_lora trains only lora adapter weights using standard training scheme.
• train_dreambooth trains a full model + identifier using dreambooth training scheme

Do I understand it correctly?

[asomoza]

yes, that's correct. Some people say that training the full model with dreambooth it's better than a lora, also it allows to later extract different loras with different ranks.

[Podidiving]

okay, now it's clear. I was under the impression that when people say about dreambooth they refer to the full-model training, but it's not always the case and it can be referred to the training strategy

thanks again!