This scene is another Phtoshop / Stable Diffusion experiment.
Unfortunately, I can’t find the image I used to guide the composition of this background (I think it was of a building under construction); I used depth-map preproccessing to control its general shape.
I wanted some sort of mechanical monster, but I didn’t want it to look like anything recognizable, like a crab or whatnot, so I drew it with pencil on paper, and took a picture.
Generating a ton of similar, but compositionally random images in Stable Diffusion is easy, but ultimately frustrating because without guidance, it will never match what you have in mind. These are outtakes from an unguided prompt “lone wanderer in ancient ruins” with several loras running for style and lighting.
I didn’t like any of them.
Here are some of the variations I made with Stable Diffusion, using the “scribble” and “lineart” Control Nets:
I had no choice but to hand-mask the monster on top of the background. Even Photoshop’s AI “select background” tool can’t figure out what’s going in fantastical images like this.
One final levels adjustment in Photoshop, then I used Ultimate Upscale SD to produce the image you see at the top of page (4160 x 2304).