I've been out of the loop when it comes to the newer lora training methods and new SD-developments in general. So any infos or tips that might be usefull would be greatly appreciated.

For now I have mainly these questions:

Is there a better program for local training than kohya ss gui? (I am so far shying away from collab and the likes, since I don't like the idea of uploading my dataset to a google server. But maybe I'm just misunderstanding that aspect of that tool so far?)

one of the greatest problems I've encountered with the results of my first training runs half a year ago were the fact that despite extensive tagging the loras would tend to replicate backgrounds, hairstyles, faces and even the general artstyle of the majority of the dataset long before the desired concept was even vaguely present. I know this is in part due to the very small size and quality of the dataset back then, but especially the style-bleed seemed to become greater the more images the dataset contained. So my question; is there a way to "force" the training to focus on the relevant part of the image other than just cropping? Because I've tried that before and then all outputs where cropped, too, and it wasn't even useable for img2img. Could training with a negative prompt help? If so, how? Is there a way to diminish the training of an artstyle other than a greater variety in styles? Would any of you more successful lora - trainers be so kind and caption one or more of the images I've posted the way you would for your datasets?

I'm sorry to bother anyone with this little wall of text, but it's not like I could expect help for this kind of lora on reddit or civit, I think.