How to train my own lora?

#1
by Vigilence - opened

I too would like to create custom loras for 3.5. With what software can I go about this?

Shakker Labs org

We use diffusers.

You could also try ai-toolkit.
The example config for SD 3.5 by Osrtis, the main maintainer of ai-toolkit, is located here.

One may set it up locally, but it is also relatively easy to run ai-toolkit configs using a Google Colab notebook.
One way to do so: There are template notebooks for Flux training at the same ai-toolkit repo, and to run an SD3.5 training one would simply need to replace some of the config parameters in one of these Flux notebooks with the settings from the above-linked example SD3.5 config. Granted, it is not very likely to work on a free-tier Colab machine (I might try to test whether it's possible later tonight). But for a $10 fee, one could run 8-9 hours of training on an A100 (and basically try any setting they like with SD3.5 training). Or potentially dozens of training hours on an L4. Though with L4, one must use relatively low rank, and smaller batch size than on an A100, and make sure to add the "lowvram: True", – or, using Colab-notebook-specific syntax: "('lowvram, True')," – to the config.

Sign up or log in to comment