@Number00

I've been playing around with it I'd say the yellow dress from the movie is in the data. Adding some combination of these booru tags to your prompt might help you:

dress, yellow dress, cleavage, collarbone, high-low skirt, back bow, sleeveless dress, single bare arm, single bare shoulder, arm ribbon, leg ribbon, green footwear

The green hair ties with the three red berries are a bit of a problem, as they will usually be the yellow bows. If you look at the training data tag JSON in the file, it includes "yellow bow". If you add that to the negative prompt, you'll be more likely to get the green. The back bow also tends to be the wrong color, and the skirt isn't too consistent either.

Anyhow, let us all know how it works out for you.

@negolah  I gave it a fast try, starting from your recommendations. Using "yellow back bow" plus your negative "yellow bow" seem to preserve for me the proper colors for the hair bow and the back bow, most of the time. With other colors (as the ribbons) it's a lottery: forcing green for the arm/leg ribbons will give you a persistent green sash for the dress.

My biggest gripe, though, was that I have been unable to get a single shot with the exact proper form for the skirt: "high-low skirt" doesn't seem the proper tag but I couldn't find anything more suitable from danbooru tags (?)

Anyway, if the generations are limited to upper body shots, I guess that good results are loable...