search

zsyzzsoft / co-mod-gan

[ICLR 2021, Spotlight] Large Scale Image Completion via Co-Modulated Generative Adversarial Networks
Other
442 stars 68 forks source link

General to arbitary image size #41

Open RyanHangZhou opened 2 years ago

RyanHangZhou commented 2 years ago

Very impressive work! Does this algorithm support image inpainting of arbitray image size? It seems to have errors if directly dealing with image size that is not 512X512.

zsyzzsoft commented 2 years ago

The model currently does not support arbitrary image size, but you can try split the image into multiple parts and inpaint each part separately if the mask is not too large.

RyanHangZhou commented 2 years ago

The model currently does not support arbitrary image size, but you can try split the image into multiple parts and inpaint each part separately if the mask is not too large.

Thanks! I tried on resized ones and it works well on small-size images. I have to say that the mask is actually very large with my setting, and I don't want to downsample it ...

zengxianyu commented 2 years ago

We cam make it generalize to arbitrary size by add a global pooling layer to the 'E_4x4' block. However I'm not sure if this would affect the performance