search

raywzy / ICT

High-Fidelity Pluralistic Image Completion with Transformers (ICCV 2021)
313 stars 44 forks source link

Pretrained weight of upsampler of Places does not work well. #36

Open UdonDa opened 2 years ago

UdonDa commented 2 years ago

Hi, @raywzy

I tried to use both FFHQ and Places2 pretrained weights of upsampler. However, the Upsampler's pretrained weight of Places does not have enough quality. We know that the weight of the generator is trained during 322000 iterations. Do you think that the attached results are correct?

From left to right, 1st stage output | blended result with masked input image | raw output of upsampler | GT image | masked input image | raw output of upsampler within given mask

debug_0

fittiing commented 1 year ago

Hello, May I ask if you encountered this error again when running this code? when I am running image

the error message shows this. How can I modify the error image