I worked on a few more variations of this series yesterday, after I devised a revised multipass workflow. Instead of using SDXL to drive SD 1.5, I tried using Turbo XL instead — which signficantly improved the speed for the first pass.
Turbo XL uses the SDXL latent, so this is not possible using Automatic1111. I plan to write a Python plugin for A1111 so I can do it inside A1111 — but until I have time to write that, these were made with ComfyUI.