It's true that image models are much less of a burden on GPU VRAM than a model like BLOOM where fitting it into a few A100s is ideal, but these diffusion models are a PITA for a ordinary hobbyist in terms of total compute: the CLIP pass over the text input is almost free, but then you feed it into the diffusion model, for one sample you'll be doing 10-100 forward passes (depending on how fancy the diffusion methods are - maybe even 1000 passes if you're using older/simpler ones), and for interactive use, you really want more like 6-9 separate samples; then they have to pass through the upscalers, which are diffusion models themselves and need to do a bunch of forward passes to denoise it. If you do 1 sample in 10s on your 1 consumer GPU, which would be pretty good, 6-9 means a joykilling minute+ wait. And then the user will pick a variation or edit one or tweak the prompt, and start all over again! It's like being back on 16kb dialup waiting for the newest .com to load.