Blurry results and non-recoverable facial features in CelebA-HQ dataset

[SherryXTChen]

Hi. Has anyone tried to reproduce the painting results on CelebA-HQ dataset? This model is trained on a randomly select subset of CelebA-HQ of size 27k for 3 days on a GPU and fine-tuned for half of a day using batch size 6. And here are some outputs I get, where the inputs come from my screenshot of the paper:

For comparison, here are the results shown in the paper (Fig 8):

For my code, the inpainted regions are blurry when the masked areas are big and the masked facial features are not recovered. I wonder if anyone has encountered similar issues or has any guess on what might be the reason. Thanks!

[SURABHI-GUPTA]

@SherryXTChen How did you train the inpainting model?

[SherryXTChen]

@SURABHI-GUPTA I was using my own perceptual loss code and there was a bug. After fixing it, I was able to follow the paper and get compatible results. Thanks.

[shangdibufashi]

@SherryXTChen Great job! Would please share your code and working model that trained using CelebA-HQ? Sadly, I don't have powerful GPU device...