Indeed, I have the freshest CUI. Also, my card is 4070ti with 12GB, so maybe that's the reason it drops dead but it is not immediately clear - especially since I did that memory tweak. Welp, maybe this is rather a torch thing, gonna be reinstalling that when the muse strikes.
Extremely that would be 30 minutes per image compared to 20 seconds for PONY or XL
I see, but iteration I thought you meant for the entire image.
What's the prettiest female the model can generate? Asking for a simplish midsized photo against a simple background.
This is using the default workflow provided with Flux, if you have a better workflow and want me to use that instead feel free to share it.
Have you tried using the launch argument --use-pytorch-cross-attention in Comfy vs default Xfomers
I haven't tried any extra arguments in Comfy yet, I'll give it a try tomorrow.
It is supposed to activate the CuDNN attention, If we get a windows build of 3 it should use CuBlass for those using an Ada Lovelace card
I don't know if the attention would affect the image quality, but did you compare your Iterations per second vs XFormers
It seems slower overall now, it was taking roughly 320 seconds for 4 images before now its taking 1200 seconds
Interesting thanks for testing, in my test I had no speed difference on any model, But I do have the CuDnn files in the Comfy Torch build.
I think for the faceswaps it's easier just to use face fusion, and the result should be good. I do love though how you made vide2video by actually making a frames animation. I was trying to do something simillar but tried only with cartoons. but my results are not stable at all. Could you give some tips for me?
