When it comes to generating images with Stable Diffusion, one of the parameters that you control is the “Sampling Steps,” which on a local installation can be anywhere from 1 to 150, whereas most online services give you a range of 10 to 100 (typically).
So the question is, are more steps better or worse? And it’s not an easy question to answer because it can vary, but in general, the more steps, the better quality your final image will be. But that doesn’t mean you should set the steps to the maximum level every time. Because it’s all about what you are trying to achieve — your end game.
Some people will tell you that you can usually get what you want with 20-25 sampling steps, and while this is true, I tend to always go much higher than that. But we’ll do a side-by-side comparison later.
If I’m testing out a new prompt, I’ll set my sampling steps at 30 to 40 so that it will process quicker, and I can see if this is even remotely what I’m looking for. Then if it does look the way I want, I’ll up my sampling steps even more and try again.
But let’s look at Charlie Monaco in a side-by-side comparison using 20, 40, 80, and 120 steps. Obviously, each time we use the exact same prompt and all other settings. The only change in each new generation is the number of sampling steps used.
I also repeated the test using Riley Evans so that you can see a great difference between the two females.
This is created using the “Lets Be Real” base model, which you can find at Civitai if you want to try these tests out yourself on your local installation.
The change is minor from steps 20 to 120, so it really depends on what you are looking for. Some may be perfectly happy with 30 or 40, others might prefer 100. As they say, beauty is in the eye of the beholder.
It really is all subjective, and that’s why the final decision on how many sampling steps to use is up to the creator. AI art really is as much art as it is science.
I thought you might like a look at Sampling from 1 to 100, image by image, step by step, with a generation of Rachel Starr.
You don’t even really good to see that it’s an image of a real woman until around step 5, but it still doesn’t look like Rachel Starr.
By the time you get to 11, it’s a much better quality image but still clearly not defined enough to be obvious it’s Rachel Starr’s face we are looking at.
By 27 you really get the full picture here, although the details in the background aren’t all that great.
By the time we get into the 50s and 60s, you really see the full picture in all its glory with subtle background details and better lighting.
So now you know how sampling steps work with stable diffusion. Or at least I hoped I helped cleared it up a bit.
