First attempt to make this glitchy FM-style LoRa. And (somehow) it works!
Just use made of lines as activation word with weight 0.8-1.
Description
First attempt.
FAQ
Comments (2)
I like this LoRA a lot, but it took a fair amount of tinkering to get it to work outside the box. If you don't force it, this LoRA wants everything to be a face. With the right strengths and prompts you can get it to admit that not every prompt should be a grayscale human face in profile.
To begin with, lower the LoRA strength to below 0.8 and increase the prompt strength to at least 1.3. It also helps to include the positive prompt "no human" and add a negative prompt for "face". Then you can really have some fun with it.
Or, of course, you could stop trying to annoy it and make lots of grayscale faces in profile. That could be fun too, I suppose.
You are absolutely right. This model was trained on a very small dataset consisting mainly of faces, that's the problem. And the lack of captions, But as I said in the description, this is only the first attempt, and I'm surprised that it generally works. Next time I will try to do better and fix mistakes.