I run it on my computer, and I think the technology is awesome. It can do things that weren't possible before. SDXL Turbo is very fast, even on cpu (20 seconds for me) and Qwen-Image can create more complex pictures (but it needs 56GB ram). All of it locally, no internet needed. I hate having to be online for stuff.
I think the technology as it is right now is "useless", as in "not productive" but I can't have a job anyway and its fun to use, kinda like a videogame (what will change if I add this word? What to add to improve the output? Should I switch to image-to-image mode?).
However, the new image-editing functionality (not the image-to-image mode) seems interesting and might even make it "useful", but my computer can't run it yet. Progress is rapid in this field, so maybe in the future.
IMO, classical image-editing programs are just the wrong approach, kinda like how vertices are the wrong approach for shadows and need tricks and workarounds while raymarching/raytracing can do shadows easily.
tips for using
CPUs can run fp32 and bf16, but not int8 or nf4, and I don't think they can do fp16.
1B parameters is 4GB in fp32 and 2GB in bf16.
The large files outside the folders are not needed for the pytorch/diffusers approach.
Make a python venv, don't install all those libraries into the system python. The 'activate' script replaces python in the current shell with the venv's python.
Get a model with "adversarial diffusion distillation" usually named "lightning" or "turbo", those are way faster. They not only use way less steps, but also don't use guidance scale, making the steps faster too.