Linaqruf / kohya-trainer

Adapted from https://note.com/kohya_ss/n/nbf7ce8d80f29 for easier cloning
Apache License 2.0
1.84k stars 302 forks source link

Revert Changes To The Stable One #232

Open Linaqruf opened 1 year ago

Linaqruf commented 1 year ago

I'm sorry, it has been 2 months since the latest update, and everything has become complicated. It makes me uncomfortable to even look at this repository. I want to revert the changes before things became complicated. Regarding the /dev/ branch, here's what I want to do:

Things I want to do:

Also, I hope you all can use Kohya LoRA Finetuning instead of Kohya LoRA Dreambooth, so I can focus on one LoRA notebook. It's basically my fault from the start to create two LoRA notebooks.

anothertal3 commented 1 year ago

Yes, I know that feeling when a code base gets out of control ;)

I haven't used the "Lora Finetuning" yet and there are a couple of settings which are new to me (e.g., "Bucketing and Latent Caching"). There also doesn't appear to be as lot annotated as in the "Lora Dreambooth" notebook. Is there some "getting started" documentation to get familiar with the notebook? Or maybe a simple "migration" guide :)?

Revert support for multi-concept training

Will it still be possible to change the dataset.toml or something similar by hand? I rarely use multi-concepts as such but I frequently use subsets to prioritize images (better images get more repetitions than those of lower quality).

merodiro commented 1 year ago

Remove the Launch Portable Web UI cell

How can we test unsupported lora types such as locon and loha without the Web UI?