I will measure the quality and time of image generation at each sampling step with Stable Diffusion.
This time, I tried generating images of supercars.
I'm using an H100 GPU. The image size is 512x512.
1 step / Time taken: 0.8 sec.
2 steps / Time taken: 1.4 sec.
5 steps / Time taken: 3.1 sec.
10 steps / Time taken: 6.0 sec.
20 steps / Time taken: 11.9 sec.
50 steps / Time taken: 29.7 sec.
100 steps / Time taken: 58.6 sec.
The computation time seems to scale linearly with the number of sampling steps.
Although images with 100 steps appear to be more detailed, there didn't seem to be much difference starting from about 20 steps.
This has made me want to try generating various images with a higher number of sampling steps.