Can the shape of output image be set to any size?

leongatys / DeepTextures

Code to synthesise textures using convolutional neural networks as described in Gatys et al. 2015 (http://arxiv.org/abs/1505.07376)

273 stars 82 forks source link

Can the shape of output image be set to any size? #11

Closed trikim closed 6 years ago

trikim commented 6 years ago

In the code, the size of the output image is the same to the input image. Can the shape of the output image be set to any size? if can, how to realize?

leongatys commented 6 years ago

In principle it is possible. But since the CNN activations change with image size due to boundary effects from 0 padding, the quality of the results is typically decreased. Thus, to generate large textures from small source images, I would recommend to generate multiple small ones. You could always initialize part of the boundary region with a previously generated texture patch so that you don't get seams at the boundary.

Hope that helps, Leon