Sampling steps are the number of times the noise is
removed during the diffusion process when generating
an image.
In other words, the more steps you have,
the more the image will be gradually improved,
resulting in a more detailed and clear result.
Role of Sampling Steps
Stable diffusion starts with complete noise and gradually
transforms it into an image.
At this time, the more sampling steps there are,
the more precisely the noise is removed, and the more
details are preserved.
If the step is too low? -> Blurry image, lack of detail
1 | 5 | 10 | 15 | 20 |
25 | 30 | 35 | 40 | 50 |
If the step is too high? -> Slow rendering, little difference
General recommendations
Low (10-20) -> When you need fast results
Normal (25-35) -> Balanced quality
High (50-70) -> When you need high quality images
ultra high quality (100+) ->No need (normally 50+ makes little difference)
60 | 70 | 80 | 90 | 100 |
Stable diffusion can use a variety of samplers,
each with a different optimal number of steps.
Sampler | Feature | Recommended Steps |
Euler a | Fast and smooth | 20~30 |
DPM ++ 2M Karras | Clear and detailed | 25~35 |
DPM ++ SDE Karras | Smooth and natural | 30~40 |
UniPC | Fast but maintains quality | 20~30 |
DDIM | Smooth but less sharp | 30~50 |
Comments
Post a Comment