This texture-synthesis technique builds on CNN and its features, and introduces a structural energy, based on correlations among features. The technique can synthesize textures with structures of various scales, local and non-local, and the combination of the two.

This novel architecture for image completion based on convolutional networks uses two auxiliary networks to learn both locally and globally consistent completion, including novel-object generation, which achieves natural completion for challenging scenes such as indoor rooms and human faces.