i get this message when the kohya ss lora starts "Replace CrossAttention.forward to use xformers"
i get this message when the kohya ss lora starts
"Replace CrossAttention.forward to use xformers"
"Replace CrossAttention.forward to use xformers"
Idoes xformers speed it up, is mine not using it?
A@Furkan Gözükara SECourses do you have any suggestions for prompts/method for inpainting face correction?
Axformers uses less vram
Ahave you checked the xformers box?
Iyea, but then i get that message so im not sure its working
Iwell its only using 10.8GB of 24 VRAM so not a big deal i guess, but if it speeds up training then id like it
Awith 24 I don't think you should use xformers.. I think it gives worse results.. mind you I only have 12 so I have never tryed myself..
Iive heard this a few times now
Ii wont worry then
Fhow much it you are getting with what batch size
Fi use original prompt. if you also upscale image with img2img and low denoise it helps significantly to fix faces
Fyou can increase batch size the significantly increase training speed
Fhowever it changes how the model trained
Fit is a machine learning topic
Alike you did in your last video? going to give that a try now..
Ado you upscale before inpaint? or is upscale usually enough ?
Fi did high res fix but img2img also works
Fi usually in paint without upscaling
Aokok.. send image to inapint, mask face, use original, and only masked area, same prompt, what level of denoise usually?
Fwell it depends on the original image quality
Fif the face is very bad
Fmake higher denoise
Fif good lower
Aok cool thanks!.. I think I might have gotten the inpainting thing to work now with my self trained model...
Fby the way you dont need special inpainting model to make this work 
Ayea I know.. but I am making really small characters and when fixing the face on something that's just a tiny part of the image a LOT of the times the head size get's wrong because I don't have any pictures where the face is such a small part of the image, then a proper inpainting models saves time because fewer of the seeds are unusable..
Fyou can also try img2img methodology
Flow denoise like 0.3
Fand 2x dimension
Fsame seed same prompt
Ayea also seems good..
Fdeepfloyd IF tutorial done
Ado you know if the open pose edior is broken?
AI get nothing when I use it..
Fhavent used recently
Iis this important for kohya ss lora training?
"A matching Triton is not available, some optimizations will not be enabled.
Error caught was: No module named 'triton'"
"A matching Triton is not available, some optimizations will not be enabled.
Error caught was: No module named 'triton'"
Fnope ignore
Ftriton not available on windows
Iah, thx
ZPlayed around with kandinsjy 2, it was ok. I still prefer auto1111
Fwhat you think about this thumbnail?
F,
ZGonna have to watch your video. Half the stuff i made in deep floyd was illegible
ZOr words were cut off
ZThumbnail looks great
Fthanks
Fhere another version
F
Fwhich one better
Fi am going to sleep hopefully see you tomorrow