Very interesting! I'd like to see an actual scooby doo show with a 3d Aesthetic like this. (Even a game)

The easiest answer is the best answer. Make two seperate images then photoshop them together that way with a slanted cut to the top layer to get that result.

If you are just having fun then instead... I've never used it personally, but I believe you'd get better results with something liked tiled diffusion region control on webui if you want to make those multi-subject images. I tried using it once and was getting all types of issues. My GPU is powerful enough that I just do it without all that but I do wish they made a "Region Prompt Control" without tiled diffusion. I'll probably try to make the thing work someday just for that reason.

And you might also try training 2 concepts on just their clothing/outfits/color scheme to add to the lora and attempt something like "xyzdaphne wearing xyzdaphneoutfit, xyzvelma wearing xyzvelmaoutfit". It also depends on the training data diversity so it doesn't force the default outfit only which is already somewhat trained the lora data you already have. Not sure how well that'll work with color control but it's worth a try if you're bored and think it may help. At the very least you can use xyzdaphneoutfit:1.5 or such without effecting the character and see if you can't force those colors in.