LINK TO THE TRAINER:
https://colab.research.google.com/drive/10652lpErVXMeNH3sEEPVsSOeH7gxsQGz
Further information:
I have been training LoRAs for a long time now since the SD 1.5 era, and, rest assured, it's not something easy to understand for newcomers. You will need to see videos, learn to caption efficiently, and learn to fix the datasets before you can even delve into training.
And, even if you do so, there are still many, MANY confusing settings with no clear explanations of what they do and what they affect.
Wouldn't you like to skip all of this hassle? Well, then you've got REALLY good news.
You don't need to do any of that anymore. With this new PONY XL LoRA trainer, you only need to download the images you want to train on and a google account. You can follow the easy, simple instructions in the notebook, and you'll be done as soon as the LoRA ends cooking! No text files, no describing, no folders, and no fidding with settings!!
However, if you want to delve in a few more technical details, I'll explain now the conceptual work scheme.
The work scheme would be as follows:
1) Before processing, it is neccesary to remove transparency and resize, so there are no surprises during training. Small images are AI-upscaled, and large images are downscaled.
2) The alpha channel is deleted from the image (if any), and an automatic tagger creates .txt files with e621 format tags without underscore.
3) The prepared dataset is moved, and LoRA training starts.
NOTE: The notebook needs the password of this journal to work. It is STRICTLY FORBIDDEN to share the link with someone so they don't need to visit this journal. If I have any notice if this happening, the notebook will be inutilized, hence, it will stop working even if you cloned it. You're warned.
NOTE 2: The notebook has many dependencies, and, over time, it is likely it will start to break up. Upon further notice, it will be neccesary for me to have at least one BRONZE badge for me to fix it in case it happens or donate to me a minimal amount of buzz yet to be determined. Otherwise, I won't hold responsible of it.