Open slashedstar opened 1 year ago
That is a big one to get as it more automates the process. From what I am reading, I would love to see it in action on SD 2.1.
Bumping this, didn't realize there was already a thing. I accidentally made a duplicate in #654 . I think it would be interesting and useful to implement GLoRA at least as a parameterization of the procedure, even without the evolutionary search piece. It's a generalization over LoRA and several other procedures. People might enjoy the flexibility of several additional new methods to play with (VPT, SSF, FacT, AdaptFormer...), which GLoRA gives you for free.
Is this ever going to be a thing? reading the paper it seems to be a nice addition!
I'd like to work on this if no one else is doing this.
I wish I had the tech ability myself but I have to say with the way Kohya works it makes a mess of SDXL due to not handling both TEs as separate entities. Onetrainer has that in their beta and it is a game changer, and add this Kohya will be left in the dust.
https://github.com/Arnav0400/ViT-Slim/tree/master/GLoRA
Since I don't have the technical knowledge to know how meaningful or relevant this is for lora in SD I'm posting it here to bring attention to it