Neural networks of suitable topology are trained with sets of images, where one image of each set depicts a garment and another pair of images of each set depicts an item of apparel from multiple viewpoints, and a final image of each set depicts a model wearing the garment and the other item of apparel. Once trained, the neural network can synthesize a new image based on input images including an image of a garment and a pair of images of another item of apparel. Quantitative parameters controlling the image synthesis permit adjustment of features of the synthetic image, including skin tone, body shape and pose of the model, as well as characteristics of the garment and other items of apparel.