-
|
I took a look at the PR (#7580) and was a bit confused as to the guidelines for usage on these new samplers. I see that the CFG used in the comparisons is 4.0, and the step counts appear to be 5-25? I tried to read the SEEDS paper for information too, but it was extremely dense lol. Paging PR author @chaObserv for help; thanks in advance! Based on how slow these samplers are, it seems that they are intended to be used with lower step counts? And what exactly is the difference between |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
|
They're singlestep samplers like dpm_2 and dpmpp_2s. A (x)S sampler performs x steps bewteen scheduled steps to try to make better solutions within intervals. SEEDS-2 and SEEDS-3 are 2S (second order) and 3S (third order) respectively. Higher order samplers are usually unstable and worse under high guidance and low steps. In my test, SEEDS-3 can be better than SEEDS-2 when using low cfg like 2.0 (without the color issue like the one in that comparison). If you have experienced other singlestep stochastic samplers (dpmpp_2S_ancestral, dpmpp_sde), it's fine to use similar settings for SEEDS. |
Beta Was this translation helpful? Give feedback.
They're singlestep samplers like dpm_2 and dpmpp_2s. A (x)S sampler performs x steps bewteen scheduled steps to try to make better solutions within intervals. SEEDS-2 and SEEDS-3 are 2S (second order) and 3S (third order) respectively. Higher order samplers are usually unstable and worse under high guidance and low steps. In my test, SEEDS-3 can be better than SEEDS-2 when using low cfg like 2.0 (without the color issue like the one in that comparison).
If you have experienced other singlestep stochastic samplers (dpmpp_2S_ancestral, dpmpp_sde), it's fine to use similar settings for SEEDS.