Spaces:
Running
You've heard of Fizgig for training loras?
Ive discovered this yesterday, Fizgig,
https://github.com/shootthesound/Fizgig
For training loras, its VERY GOOD... and it takes a lot less VRAM than other software and the likeness of the person comes much faster with fewer epochs. It can run on 12gb-16gb vram gpus. You can train Minimax, Krea2, probably anything you want to.
I just thought people should know lol
nothin
Malcolm replied to a reddit post of mine where I had suggested the same thing and here's his reply:
"I am trying Fizgig and the results are better but not great still.
I am able to use the lora trained and reference images to make the resemblance even better than just the reference images (especially for someone who is not in the model; for those that are in the model only a few images do the trick).
Unfortunately, current training methods still bleed over. Even though I can make a great character using lora and images, once I put more characters in the scene - the other ones are affected by the lora. (I have not tested AI Toolkit yet, it had the way to isolate a character and maybe it would work here)"