Abstract: In the present paper, we propose a new approach for realistic image noise modeling based on a generative adversarial network (GAN). The model aims to boost performance of a deep network denoiser for real-world denoising. Although deep network denoisers, such as a denoising convolutional neural network, can achieve state-of-the-art denoised results on synthetic noise, they perform poorly on real-world noisy images. To address this, we propose a two-step model. First, the images are converted to raw image data before adding noise. We then trained a GAN to estimate the noise distribution over a large collection of images (1 million). The estimated noise was used to train a deep neural network denoiser. Extensive experiments demonstrated that our new noise model achieves state-of-the-art performance on real raw images from the Smartphone Image Denoising Dataset benchmark.