Do you use the left over noise option? In the kssmpler advance if you pick the left over noise it would send de image without denoising to the next ksampler. Also try without send left over noise and inject noise after each ksampler separately with a noise node. it add more details those 2 options should improve the 3 ksampler setup and give each ksampler the option to fix errors in the composition
Thanks. But IMHO give credit where credit is due, this method was popularized by Olivio Sarikas. Also, image composition might change, since you are first doing full denoising in 10 steps. If you did it with 3-4 steps, you wouldn't even get a proper image. So you are basically not letting the model figure out all the details it might do with that noise seed, if you use low step count in this "first phase". So this 1st pass basically makes generation to lock into certain main shapes early on with this denoising "strategy", and then the following passes will refine the shapes.
Would be awesome if you could further optimise this technique with say Impact nodes cfgschedulerhookprovider and stepsschedulerHookProvider nodes. I havent checked though if they work with flux, but they should
Do you use the left over noise option? In the kssmpler advance if you pick the left over noise it would send de image without denoising to the next ksampler. Also try without send left over noise and inject noise after each ksampler separately with a noise node. it add more details those 2 options should improve the 3 ksampler setup and give each ksampler the option to fix errors in the composition
Thank you so much. Stunning quality with Pixelwave 🤩
I tried your workflow with the PixelWave model. Not good results. Do I need to tweak the guidance value? Any others ideas ? Thx
Thanks. But IMHO give credit where credit is due, this method was popularized by Olivio Sarikas. Also, image composition might change, since you are first doing full denoising in 10 steps. If you did it with 3-4 steps, you wouldn't even get a proper image. So you are basically not letting the model figure out all the details it might do with that noise seed, if you use low step count in this "first phase". So this 1st pass basically makes generation to lock into certain main shapes early on with this denoising "strategy", and then the following passes will refine the shapes.
ComfyUI makes my head hurt! But, every time I see a video like this i have to try again. {sigh} God help me...
Just one node at a time😊
Would be awesome if you could further optimise this technique with say Impact nodes cfgschedulerhookprovider and stepsschedulerHookProvider nodes. I havent checked though if they work with flux, but they should