I'm not a good artist by any stretch of the imagination, but I went ahead and trained on my own hand drawn sketches. The results were genuinely interesting and a little bit useful too.

It does a good job conforming to my style, it's really fascinating seeing the AI output something like what I'd draw. Also, despite the results often being aesthetically unpleasing it's very different from the norm, so it's good to see it's flexible enough to handle this kind of hyper-minimalist style. I also kinda like it, but maybe that's just me.

Now you might be asking yourself, why would I want a LoRA that makes my resulting generations worse?

One of the initial use cases I had for the model, which it performs decently at, is generating a starting point base image that can be developed into something genuinely good. Give it a simple prompt and it'll make a minimalist black and white sketch that you can easily sketch on, color in, and use img2img and inpainting to improve iteratively. Oh hey, go read my articles. Similarly, it can be useful in the early stages of developing sketches, making it somewhat easier to make a really rough sketch into something cohesive.

Another use case I discovered while playing around with it is that the hyper-simplification effect forces models to show you what it really cares about, so it makes a great diagnostic tool. That is, you can ask "What did this LoRA really learn?" and literally see it. An important note for this use case is that you can keep repeating the trigger word "urusan drawing" over and over to get more and more simplification, which allows you to push it to whatever level of simplification you'd like to view your model at.

One last thing is that it makes a decent negative prompt LoRA. It reduces jaggy lines and oversimplified details when used this way. That said, I've only gotten modest results from this and I think it's mostly because the model isn't bad enough, but maybe I'm just not using it right.

The main trigger word is "urusan drawing". Repeat this separated by commas as many times as needed to get the desired amount of simplification (and my hand drawn style).

It's been trained on a few other captions as well, the dataset is included in case you're curious (note that some of the dataset is mildly NSFW).

There's some substantial style contamination from my oatd (and galena bikini) regularization datasets, so when the simplification is intermediate you'll probably see the oatd style shine through.