Upscaling in a CNN: Conv Transpose or Tile(Nearest neighbor)?

Question

So far I've seen upscaling in a net using conv transpose (for example in the DCGAN paper).

Now I'm reading a new article by Nvidia (growing GANs- see https://arxiv.org/abs/1710.10196) where they are using tf.tile to "upscale",after which they are using regular convolutions (same padding).

Where can I read more about Nvidia's researchers approach? What is the trade-off between those two approaches?

Answer 1

I would recommend this Distill article which talks in more details about using "deconvolution" layers vs "upscale + normal convolution".

In particular it seems the latter introduces less artifacts in the generated images.