Machida is a minor character who as of the upload date, has appeared in 8 chapters of the Don't Toy With Me, Miss Nagatoro manga. No official colored illustrations of her exist. No fanart of her exists. She appears in ~3 dozen black-and-white panels, most in extremely low resolution. This LoRA was a challenge to see if I could make a decent LoRA of such a character.
Overall, I'm not quite satisfied. I'll continue to work on this a bit further, but I do think it's feasible to get a good LoRA of minor manga-only characters with enough effort. This opens up a lot of new possibilities. This was the most fun I had making a LoRA, as it was a huge technical challenge.
Use at weight 0.5 - 0.9.
Keyword is machida. Prompt (black_dress, choker, gothic) for her black dress.
Issues
Again, this was a proof-of-concept LoRA, so don't expect the greatest LoRA out there.
Some details aren't quite captured. For example, her "fuzzy" irises and semi-squinting eyes were difficult to train. You'll need multiple generations to get good pics.
Why are her eyes light green? Because reality can be whatever I want.
Pics 1-2: YesMix
Description
Trained on 28 images.
FAQ
Comments (3)
I'm intrigued by your process to perfect a "low amount of material", manga-only character lora! If you come up with a new process for this, it could prove a breakthrough for characters who have a handful of official art references or less.
Lots and lots of ControlNet + Photoshop. For example, you can get a pretty decent reference image from a low-res manga panel if you use canny-style ControlNet.
Here is one panel: https://i.imgur.com/K3vLtoP.png
And here is the same panel after ControlNet + Photoshop: https://i.imgur.com/itqmnQU.png
Another trick is creating an intermediate LoRA to generate some new images and using those images in the training set of the final LoRA. I was able to get ~30 color images of Machida that way
Wow! Amazing work - thanks for the upload!