Feature Request - Choosing CLIP tokenizer / text encoder during merge

[Luke2642]

In dreambooth training, you can choose to also train the text encoder, or not.

It could be useful to be able to choose whether or not the text encoder is merged when you merge two models.

This could be implemented as a checkbox, or as a separate slider?

Maybe the other components, the VAE and U-net (that make up a stable diffusion model) could be selectively merged too.

Or, perhaps merging could be given coarse/fine features selection, like when choosing which layers are affected by a hypernetwork?

[Luke2642]

Replying to myself here, it appears a commit yesterday fixed a bug around text encoder merging for add difference, which is kinda one step closer to this idea!

#6728

[Luke2642]

It seems this idea has been largely implemented in an extension!

https://github.com/bbc-mc/sdweb-merge-block-weighted-gui