Semantic Pyramid for Image Generation

Semantic Pyramid for Image Generation. We introduce a new image generative model that is designed and trained to leverage the hierarchical space of deep-features learned by a pre-trained classification network. Our model provides a unified versatile framework for various image generation and manipulation tasks, including: (a) generating images with a controllable extent of semantic similarity to a reference image, obtained by reconstructing images from different layers of a classification model; (b) generating realistic image samples from unnatural reference image such as line drawings; (c) semantically compositing different images, and (d) controlling the semantic content of an image by enforcing a new, modified class label.