biased training data
#1
by PatKim - opened
Hello. I've tried using the model you shared. For photorealistic images, the details come to life while the skin tends to appear slightly smoother. When inputting 2D anime images, there's a phenomenon where it changes from 2D to real. I suspect this is because it was trained solely on photorealistic training data. Could that be correct..?
Agreed! And I've noticed that when I upscale images without people, it tends to add people
Thanks for your kind check.
Is there any plan to improve this in the next version?
Happy New Year!