High resolution

Can the network be trained so that the features of the style image (for example, large brush strokes) be applied to content images that are larger than the trained images?

For example, if I train the network on the 250px dataset using a style image of 750px, when I evaluate/style a content image that is 1500px all of the style image features are very small. I would ideally like the features to be of a similar size in relation to the style image.

Is this possible?

This method could be a solution: https://github.com/LouieYang/stroke-controllable-fast-style-transfer