In this work, we show the generative capability of an image classifier\nnetwork by synthesizing high-resolution, photo-realistic, and diverse images at\nscale. The overall methodology, called Synthesize-It-Classifier (STIC), does\nnot require an explicit generator network to estimate the density of the data\ndistribution and sample images from that, but instead uses the classifier's\nknowledge of the boundary to perform gradient ascent w.r.t. class logits and\nthen synthesizes images using Gram Matrix Metropolis Adjusted Langevin\nAlgorithm (GRMALA) by drawing on a blank canvas. During training, the\nclassifier iteratively uses these synthesized images as fake samples and\nre-estimates the class boundary in a recurrent fashion to improve both the\nclassification accuracy and quality of synthetic images. The STIC shows the\nmixing of the hard fake samples (i.e. those synthesized by the one hot class\nconditioning), and the soft fake samples (which are synthesized as a convex\ncombination of classes, i.e. a mixup of classes) improves class interpolation.\nWe demonstrate an Attentive-STIC network that shows an iterative drawing of\nsynthesized images on the ImageNet dataset that has thousands of classes. In\naddition, we introduce the synthesis using a class conditional score classifier\n(Score-STIC) instead of a normal image classifier and show improved results on\nseveral real-world datasets, i.e. ImageNet, LSUN, and CIFAR 10.\n