I would generate one image, say background first. Write the generator to have quite specific conditions I.e. depth of view, positioning, scale, size etc etc (I’m totally clueless as to the limitations here so again please correct if not possible). Result, one image. Backdrop.

Images can be fine-tuned with tags (albeit the AI tries to), so you can change the conditions of the images.

Then generate the character image. Maybe whole separate generator. Again, very specific with scale or pose, positioning etc and without background.

You can also do the character in the same generator, no need to code another one, you just need to change the prompts.

My thoughts were the locked-layer-combination!? To combine the two images 🤷🏻‍♂️ and voila.

The layering part is quite easy to implement (see this example generator I made a while back) but you would need to edit the character image to be cropped so it can overlay successfully on top of the image. If you were to use 'inpainting' you could just 'paint-over' the current image without needing to layer and position multiple images.

The next step would be to then adjust parameters within each generator but revolving around the same seed (might be talking nonsense here as again, I’m just learning) to adjust in the desired way but retaining the features I.e. colours, style etc for continuity.

Using the same seed is a good idea to have the 'same' essence, or replicable composition (see the prompting guide that I linked earlier). But yeah, a local setup with 'inpainting' would be better, there are also other 'plugins' for the local setups to instruct the AI area by area. Also note that you might need a powerful computer (and graphics card) to run the local models quickly.