Model training - an open secret?!
Well, some of you would say: "It's easy and self-explanatory." ...
Yes, that's true. It's easy and almost self-explanatory to create a "model" ... for example known as LoRA.
The open secret is: how do you train a model called a "checkpoint"? I've reviewed hundreds of pages and spent a lot of time trying to find a solution to this problem.
The good thing is: I found a way to do it. Currently, I'm doing training for my own checkpoint. If the setup works properly, I'll share the configuration with you because I don't like the way other people have handled it.
I don't like the idea of paying for such information and I don't like answers that are wrong or incomplete, leaving you unable to create one yourself anyway.
For the critical people among you: I'm not talking about "merging" a LoRA into checkpoint, but about the actual training.
So if you're still looking for a solution, you can just check my profile from time to time.
For anyone who wants to try it out for themselves before I finish my configuration, this article helped me get started:
https://civitai.com/articles/7615/flux-dreambooth-finetune-settings-kohyass
but as the user himself mentioned, his configuration is not working properly yet. Nevertheless, I checked his configuration and tweaked a few points in it to make it a working configuration.
So if anyone else has helpful information on this, feel free to share your information as well.
Nevertheless, I will publish my configuration within the first checkpoint I can create properly. Since my PC is not the best, it will take some time to train a checkpoint... and then I have to evaluate its functionality. The first two iterations trained on Kohya_ss as expected, but ComfyUI could no longer identify the file as a checkpoint… So some more testing is needed.
Stay tuned ;)
