We present a generative adversarial network-based system that faithfully colorizes black and white images without human intervention. Recent methods for such problems typically use per-pixel loss between the output and ground-truth images and the images generated using such loss lacks details. where as in "Perceptual losses for real-time style transfer and super- resolution" [2] paper suggests the use of the perceptual loss function that generates high-quality images. We combine the benefits of both approaches. we have replaced pixel-loss function with perceptual loss function which gives visually pleasing results and used discriminative learning technique where we train first half of generator with lower learning rate as we are using pre-trained model and last half with higher learning rate which reduces the training time of generator. It’s easier and faster to train a large number of samples of smaller images initially and then scale up the network by improving images to 220px by 220px from 64px by 64px. This is called progressive resizing [6]. it also helps the model to generalize better as is sees many more images of different shapes and less likely to be overfitting.