Getting to run JoePenna / Dreambooth-Stable-Diffusion in a 16GB Google Colab

[jslegers]

I've currently been using @ShivamShrirao's fork of huggingface/diffusers (ShivamShrirao/diffusers) and I've been running this perfectly fine with my own finetuned Google Colab notebook.

Is there someone who can tweak this repo so it can run on 16GB VRAM as well? Or, since I'm especially interested in @kanewallmann's addition for support for multiple concepts & classes (kanewallmann/Dreambooth-Stable-Diffusion, is there someone who can port that feature to huggingface/diffusers (ShivamShrirao/diffusers)?

Related :

• Support for multiple fine-tuned concepts & classes on one model with Dreambooth ShivamShrirao/diffusers#21
• Support for multiple fine-tuned concepts & classes on one model with Dreambooth huggingface/diffusers#752

[djbielejeski]

Shivam's repo is diffuser based, not sure what it would take to get some of those optimizations in this repo.

[jslegers]

@djbielejeski :

So you just close the issue as "completed"? That's not very helpful / useful...

So, I guess this version of Dreambooth is not diffusers based? Could you enlighten us on how it is implemented, then? Maybe this could shed a light on potential optimisations...