I SUCK at image gen.

So I started out wanting to do an illustration for my Tama 9th district side story. Figured I'd throw the 2 magical girls together and call it a day. Easy enough. When I was writing, I didn't really care much for the specifics, so I ended up with these descriptions and additions:

Starlight Spice/Karen: "She wore a frilly miniskirt and matching bolero over a crop-top, with mismatched splashes of yellow, purple, and pink that created a pastel riot on her outfit." I decided to give her dark purple hair to contrast the bright colours, and make her smug to go with her confident attitude.

Pretty Kitty/Quinn: "The girl's outfit was all bright stripes and block colours, with a ridiculous-looking pair of cat ears on her headband." She also had a cell phone. I decided to give her a pouty vibe and dark primary colours to contrast with the bright stripes. The orange drill hair was an afterthought, since I made Karen's art first.

Phase 1:

Starlight Spice

Karen

Pretty Kitty

Quinn

A suitable background

Tama Alley

Phase 2:

I stuck them in paint.net and used the Magic Brush to try and remove the white backgrounds. Then, I resized a bit to fit them all in the same image.
The result was disappointing. I think it's easiest to see around Karen's purple hair.

District Dorks

Phase 3:

Decided to do what anon suggested, mess around with img2img and use the 2 earlier pictures as vibe transfer fuel. Also re-read my story and realized that Starlight Spice was meant to have a crop-top, so I added it to the description a bit while vibing (I liked the previous design better though, tbh).

Basically started with a simple: 2girls, outdoors, magical girl, side-by-side
next, vibed the left girl generated, added Starlight Spice's descriptions (stuff like: smug, yellow cropped jacket, frilled skirt, purple hair, etc) and the vibe image
finally, vibed the right girl generated, added Pretty Kitty descriptions (stuff like: bored, orange drill hair, cell phone, cat ear headband, etc) and the vibe image

Better?

District Dorks 2.0

Ending thoughts

I still hate generating art of multiple characters, but I guess it worked a lot better. If there's a better workflow than the one I described, I'm all ears.