To this:
But the process is different for removing noise depending on which schedule type you use:
In this section, I am going to cover all the noise scheduling methods in WebUi.
This will give you a background of how each one works before I tell you the best ones to use in the next section.
Karras follows an algorithm that generates noise based on the discretizations of Euclidian distance.
To be honest, I don’t know what that means.
But what matters is that Karras adjusts the noise scheduling depending on the number of steps. It will remove noise at progressively smaller steps at the end of the process.
It is a lot more flexible than other fixed schedulers, making it very reliable.
Exponential is exactly the same as we learned it in school.
It removes noise very rapidly at the beginning steps but reduces the amount of noise removed later.
It is best for using a small amount of steps per image.
Uniform reduces noise at a consistent amount per image.
It can have some non optimal performance compared to other types due to it not being able to focus on critical steps in the process.
Polyexponential uses a combination of exponential rates.
This enables the model to be able to adapt to the data distribution better.
SGMs estimate a gradient or score based on the noise level data to add and remove noise.
TLDR: Karras is the best.
The generally recommended schedule type is Karras. This is because of Its ability to adjust noise removal using its formula makes it the best.
Secondly, I wouldn’t recommend both uniform types as much as the other. Especially just uniform.
Below, I made a x/y/z plot of the schedule types from above.
I am going to compare the quality of images created with different sampling steps amounts to see which one has the highest quality:
Prompt: race car on a icy mountain peak, snowy landscape, dramatic lighting, cinematic, establishing shot, extremely high detail, photorealistic
Most of the images look similar but uniform has worst images at lower and higher steps.
Otherwise, the three schedulers in the middle have the best looking images with karras looking the best.
There you have it.
Everything you need to know about stable diffusion schedule types.
Let me know if I missed anything or add something you want to see by leaving a comment below.