The background was generated from a photo I took at a local farmer’s market. The character to the left was composited using depth map pre-processing. Note that Stable Diffusion tends to repeat faces, but that’s fine for concept art if they’re just background.
This is just one other version I generated without upscaling, there’s slightly different poses and clothing on the people, but generally the scene composition is the same.
The trick to making things deliberately with AI tools is to prepare your source materials appropriately and manage the “freedom” of the AI by using ControlNet plugins - namely OpenPose, Depthmaps, and Canny.