NM-FlowGAN: Pixel-wise noise and spatial correlation modeling for sRGB noise without paired images in generation time

sRGB noise modeling refers to replicating camera noise in the standard RGB color space, which is critical for training robust image denoising networks when collecting real noisy-clean pairs is difficult or impossible. Accurate modeling of real sRGB noise requires capturing two key components: pixel-wise noise characteristics and spatial correlation across neighboring pixels. Pixel-wise noise depends on a multitude of factors such as clean image intensity, camera type, and ISO settings, while spatial correlation arises from complex image signal processing (e.g., demosaicing) that induces inter-pixel dependencies. To effectively model both the pixel-wise noise and the spatial correlation, we propose NM-FlowGAN, specifically designed to model both pixel-wise noise and spatial correlation in real sRGB images. In particular, our pixel-wise noise modeling network, based on Normalizing Flows, exploits its high training stability to learn noise characteristics affected by multiple factors, while our GAN-based spatial correlation network efficiently captures pixel-to-pixel relationships. Furthermore, in contrast to recently proposed methods that rely on real paired noisy images in generation time, our approach requires only clean images and easily obtainable camera parameters, making it widely applicable in scenarios where collecting noisy-clean pairs is impractical. Experimental results demonstrate that NM-FlowGAN not only outperforms other baselines in sRGB noise synthesis but also yields superior denoising performance when the synthesized image pairs are used to train denoising networks.