Replies: 3 comments 7 replies
-
i think these two ways is totally different... in my opinion,the second KSample will not continue the work from the first one...because that the KSample has no pause/continue state every time start a KSample,it will start from step0,buy in your case, the second one's input params is the first one's output,and the first one's input is text description,for that,the finally result is just the result of KSample2,not KSample2 continue from KSample1 so step14+step14 will no same as step28.. but i am also new here,the above is just my personal understanding,maybe its wrong |
Beta Was this translation helpful? Give feedback.
-
After testing, I get it. Let me explain it for you. First, the 14-step KSampler and the 28-step KSampler are totally different from each other. Thanks to Raveness:
The 14-step KSampler divides the work into 14 pieces while the 28-step KSampler divides the work into 28 smaller pieces. So their works can not replace each other. So you need two 28-step KSamplers, each one does half of the work. the fomula In short: Second, you should be careful with the add/clear noise. For the two KS in series, the first KS should enable Think of it like this workflow:
if you do not enable
that will cause the result a bit of different. For more KSs in series, make sure only the first one adds noise and only the last one clears leftover noise. Here is the workflow. |
Beta Was this translation helpful? Give feedback.
-
BTW, according to our conclusion, this picture is WRONG. source : https://imgur.com/a/VjFi5uM It's NOT the way how the AI draw. Because the way to generate this X/Y/Z plot by au111 is like:
It means that to generate the whole image needs 3+5+7+10 = 25 iterations totally. But according to our conlusion1, n-step KSs cannot replace each other when n is varied. The real way how AI draw is like (in ComfyUI): And it only takes 10 iterations! (vs. 3+5+7+10 = 25 iterations in au111) here is the workflow, you can test it yourself. Another reason to turn to ComfyUI. |
Beta Was this translation helpful? Give feedback.
-
I used two 14-step KSamplers in series to generate the left one, and used a 28-step KSampler for the right one. But they are obviously different. But why? I think
14 steps + 14 steps = 28 steps
so they should be the same result.My workflow:
ComfyUI_00045_.zip
(The model named after
nai
is the leak novelai model)Imagination result:
Beta Was this translation helpful? Give feedback.
All reactions