[Feature Request]: Possible storage saving through nondestructive dehydration of model/checkpoint

[Centurion-Rome]

trafficstars

Is there an existing issue for this?

• [X] I have searched the existing issues and checked the recent builds/commits

What would your feature do ?

Possible memory storage saving

Proposed workflow

Integration of a tool that allows the "dehydration" of a Dreamboot model in a nondestructive way so that it can fit in 1GB of space: (https://github.com/bmaltais/dehydrate)

Additional information

More details: https://www.reddit.com/r/StableDiffusion/comments/yg0k6r/dehydrate_dreambooth_model_so_save_space_and/

[Thomas-MMJ]

It doesn't allow memory savings, it allows disk storage savings.

[Kavukamari]

I'm wondering if dehydrating a model and then adding it to another model is the same as training the "another model" on the data from the dehydrated portion, because if that's the case, we could simply load Stable Diffusion as a base, and just add dehydrated modules on top of it dynamically while rendering

[Thomas-MMJ]

I'm wondering if dehydrating a model and then adding it to another model is the same as training the "another model" on the data from the dehydrated portion, because if that's the case, we could simply load Stable Diffusion as a base, and just add dehydrated modules on top of it dynamically while rendering

Essentially yes, it is the weights that were changed when training the new model on the base model. You probably wouldn't want to change it during rendering but it could be changed during loading.