NAFNet is a simple baseline for image restoration that removes nonlinear activation functions from the network architecture, relying instead on linear operations such as convolutions and normalization layers. The approach demonstrates that nonlinear activations are not essential for image restoration tasks and can be replaced with simpler building blocks while maintaining or exceeding performance. It achieves state-of-the-art results on benchmarks including GoPro for deblurring, SIDD for denoising, and stereo super-resolution datasets.