This composite did not start as a finished image.

I generated a simple, non photorealistic base.
A watercolor style interior of a train.

From there, I did not ask the model to generate a new picture.

I generated one element at a time.
Each element was created to sit on top of the previous one.
Each layer became the base for the next.

By generating incrementally instead of regenerating the whole image,
I could treat the result like a traditional composite.
I could merge layers.
Adjust colour.
Address matting issues.
Fix edges.