Thanks a lot for the great efforts! Although the Samsung paper didn'

Bilinear Down/Upsampling about realistic-neural-talking-head-models HOT 5 CLOSED

vincent-thevenin commented on July 19, 2024

Bilinear Down/Upsampling

from realistic-neural-talking-head-models.

Comments (5)

Connor323 commented on July 19, 2024 1

This is the checkerboard effect I got using default nearest interpolation:

This is the ringing effect I got using bilinear interpolation:

from realistic-neural-talking-head-models.

vincent-thevenin commented on July 19, 2024 1

Looks like the ringing effect is somewhat less noticeable than the checkerboard one. If you'd like to do a pull request go ahead :)
The effect should die out eventually with more training.

from realistic-neural-talking-head-models.

mrgloom commented on July 19, 2024

Looks like nn.Upsample use nearest upsampling algorithm by default.
https://pytorch.org/docs/stable/_modules/torch/nn/modules/upsampling.html

Realistic-Neural-Talking-Head-Models/network/blocks.py

Line 153 in 0c9b470

self.upsample = nn.Upsample(size = out_size, scale_factor=scale)

from realistic-neural-talking-head-models.

Connor323 commented on July 19, 2024

Correct! And I have been testing changing it to bilinear interpolation. It seems it removes the checkerboard effect, but introduces a ringing effect. lol I'm still working on it.

from realistic-neural-talking-head-models.

mrgloom commented on July 19, 2024

What do you mean by 'ringing effect'? can you show example image?

from realistic-neural-talking-head-models.

Recommend Projects