On intel MacBookPro 2020, CPU-only, the original one[1] using pytorch utilized one core only. A tensorflow implementation[2] with oneDNN support which utilized most of the cores ran at ~11sec/iteration. Another OpenVINO based implementation[3] ran at ~6.0sec/iteration.

[1] https://github.com/CompVis/stable-diffusion/

[2] https://github.com/divamgupta/stable-diffusion-tensorflow/

[3] https://github.com/bes-dev/stable_diffusion.openvino/

Yes, I use [3] and I get 2.4s/iter on my 10 core machine. I was wondering if keras would give additional help here. I'll have to try I guess.

tried it yesterday, on intel i9 macbook pro it takes about 300 seconds per image.

You mean the keras version? How does it compare to the original one? Currently on my 10850k I get 2.4s/iteration, which is borderline usable. I haven't managed (nor tried very hard) to get the cuda version working on my 1070; I expect to be a little better, but I don't want to fight with ram issues.

How many steps did you perform?

I tried some and found no major differences after 16 steps or so with given random seed.