Conditional Generative ConvNets for Exemplar-Based Texture Synthesis

The goal of exemplar-based texture synthesis is to generate images that are visually similar a given exemplar. Recently, promising results have been reported by methods relying on convolutional neural networks (ConvNets) pretrained large-scale image datasets. However, these difficulties in synthesizing textures with non-local structures and extending dynamic or sound textures. In this article, we present conditional generative ConvNet (cgCNN) model which combines deep statistics the probabilistic framework (gCNN) model. Given exemplar, cgCNN defines distribution using ConvNet, synthesizes new sampling from distribution. contrast previous models, proposed does not rely pre-trained ConvNets but learns weights for each input exemplar instead. As result, can synthesize high quality dynamic, unified manner. We also explore theoretical connections between our other models. Further investigations show be easily generalized expansion inpainting. Extensive experiments demonstrate achieve better at least comparable than state-of-the-art methods.