Verifying the conribution of two compoents of our proposed solution.
Multi scale discriminator.
Reconstruction loss.
Results show (left to right) retargeting with single scale discriminator (Single scale D),
retargeting with InGAN trained without reconstruction loss (No reconstruct),
finally, we show results of InGAN with all components (InGAN).
See Sec. 5.3 in the paper for more details.
Birds
input image
Note how using only single scale dicriminator (right) produce result that has no global structure only very local textures are preserved.
When discarding the reconstruction loss (middle) the model experience a "mode collapse" - only the dominant bird is kept while the other two are discarded.
In contrast, using multi-scale discriminator and reconstruction loss keeps all three birds (right).
Wdith of original image was changed in range x0.4 to x1.75.
