Sign In

Basic tips on wrangling manga (and comic) data sets

Basic tips on wrangling manga (and comic) data sets

When I initially wrote this I had done 4 LoRAs with significant parts of the data set made up of manga (2 entirely so), and I don’t really see instructions on this covered in guides (most expect anime/game screenshots or booru scrapes to be the main data), so I made an article for it. Since making this article I've made 4 more LoRAs based entirely off training data from American comics and 1 based on a old newspaper comic using largely the same techniques and can safely say it is entirely doable to make a character LoRA from comics alone. Please give any corrections or additions.

For general LoRA making instructions, see this article.

Cropping

To start, you're really are just taking pictures pages and cropping out the character then saving the result. I generally favor cropping slightly narrower than you would normally to avoid another character overlapping, though this isn't always doable. If I have to use such pics, I prefer to include enough of the other character it can be easily recognized as (part of) another person, so I can tag the image as having multiple people and hope the training can understand that, rather than hope it can filter out stray elbows it doesn't really know are elbows.

Blank Out Speech Bubbles

This is pretty simple: If you have to include a speech bubble (in whole or in part), remove all the text within it. This is a simple process of just deleting (if your image editor defaults/ is set to plain white for deleting) or covering it with (a) white box(es). Text really confuses SD, but speech bubbles aren’t nearly as bad. Once you’ve done this, tag all applicable images “speech bubble, blank speech bubble”.

Clear out extraneous stuff from blank backgrounds

Blank backgrounds are an opportunity simply delete stray shoulders, sound effects, panel dividers etc. from an image. If you can't get all of a object out, don't get any of it, but it's often possible to clean up an image with minimal fuss. If you're feeling really fancy on more complex backgrounds you can try cutting out just the character you want in GIMP (or some other image manipulation program) but I wouldn't bother unless you really need pictures in your data set and have a bunch of other characters in view.

Tag for Negative Use:

You should be doing this with anything, but it’s especially vital here. Start with tagging all your manga images “monochrome, greyscale, manga illustration” (unless it’s got a cover/special page that’s in color of course, then just manga illustration) and images with comic effects like panel divides as "comic". This isn’t so you can prompt it (though you can) but for use as a negative. PDXL handles such style negatives relatively well (I think it even handles them being absent from the prompt well in most cases), but a LoRA needs to know these are a style. Also worth learning and using tags seen on "low detail" shots like "dot eyes" or "chibi" as well as things like "sound effects" or "emphasis lines" so they can be applied.

Don’t Tag Colors (except black/white/grey/dark/light):

Even if you know a color is supposed to be, it’s not “actually” that color in black and white. PDXL can insert colors well enough at prompt time. Even without specifying colors and using a data set devoid of any colors (such as with my Mana 2.0 LoRA) PDXL is actually quite good at inserting a fitting color when "monochrome, greyscale" is in the negative.

Edit (October 16th 2024):
A few updates. Added section on cropping, noted the ability to cut a character out in GIMP and a bit more commentary on colors.

Edit (Nov 7): Minor copy editing. Updated intro to reflect my further experience with this type of LoRA.

17

Comments