idream -getting poor results
orydatadudes opened this issue · comments
orydatadudes commented
HI
thank you for idream implementation
i used DeepFashion dataset to check if i can use change human pose ( like the paper show )
unfortunately i didn't succeed to get good generated images
for example :
source image
after trying to change posture (“A person with crossed arms”)
i got bad results ( for different levels of alpha )
i tried with more images and more text prompt and increase alpha range but nothing works for me
did someone succeed getting good results ?
thanks
Justin commented
Yeah it seems quite sensitive to parameters, I found sometimes that doing the first stage training, (training the embedding) for longer helps.