All valid points, of course. As an independent explorer I adapted my workflow to use night's worth of workstation compute to generate a crop of new images from a simple templated prompt language. It also helps a lot to have at least two presets - "exploratory" and "hq", to minimize iteration time and maximize quality of promising prompts.

Still, I think optimization of diffusion models for efficient inference isn't yet pushed to the limits. At least if we look at what's available to the public - AFAIK public inference software distributions didn't even quantize their weights.