Fig. 1.

Image synthesis algorithm and example synths. (A) Schematic of deep image synthesis algorithm. We pass a natural image into an Imagenet-trained VGG19 model and extract intermediate layer activations from layers pool1, pool2, and pool4. Then, we compute the correlation between pairs of feature maps within a layer (Gramian) constrained within spatial pooling regions (1 × 1, 2 × 2, 3 × 3, or 4 × 4). To synthesize, we iteratively update the pixels of a random seed image using gradient descent to match the spatially pooled Gramians from the original image. (B) Example natural images and feature-matched synths, varying in feature complexity (columns 2 to 4, fixed at 1 × 1 spatial constraint), and varying in spatial constraint (columns 5 to 8, fixed at pool4 complexity).