FFDNet is a fast and flexible convolutional neural network (CNN) designed for image denoising. It addresses the limitations of existing methods by using a tunable noise level map as input, allowing it to handle a wide range of noise levels and spatially variant noise. Unlike previous discriminative denoisers, FFDNet can effectively manage different noise levels with a single network, and it is faster than BM3D even on CPU without sacrificing denoising performance. The network processes downsampled sub-images, achieving a good balance between inference speed and denoising performance. FFDNet is evaluated on both synthetic and real noisy images, showing superior performance in terms of denoising quality and computational efficiency. It performs well on images corrupted by spatially variant AWGN and real-world noisy images with signal-dependent, non-Gaussian noise. The model is trained on a large dataset of input-output pairs, and it is tested on various datasets including BSD68, Set12, CBSD68, Kodak24, McMaster, and RNI6. FFDNet outperforms state-of-the-art denoisers such as BM3D, WNNM, MLP, TNRD, and DnCNN in terms of PSNR and visual quality. It is also effective in handling spatially variant noise by specifying a non-uniform noise level map. The model is flexible and efficient, making it suitable for practical denoising applications. FFDNet is trained with a noise level map and uses orthogonal initialization to avoid visual artifacts. It is compared with other methods in terms of noise level sensitivity and real noisy image denoising, showing robust performance across different noise levels and image structures. The results demonstrate that FFDNet is a highly effective and efficient solution for image denoising.FFDNet is a fast and flexible convolutional neural network (CNN) designed for image denoising. It addresses the limitations of existing methods by using a tunable noise level map as input, allowing it to handle a wide range of noise levels and spatially variant noise. Unlike previous discriminative denoisers, FFDNet can effectively manage different noise levels with a single network, and it is faster than BM3D even on CPU without sacrificing denoising performance. The network processes downsampled sub-images, achieving a good balance between inference speed and denoising performance. FFDNet is evaluated on both synthetic and real noisy images, showing superior performance in terms of denoising quality and computational efficiency. It performs well on images corrupted by spatially variant AWGN and real-world noisy images with signal-dependent, non-Gaussian noise. The model is trained on a large dataset of input-output pairs, and it is tested on various datasets including BSD68, Set12, CBSD68, Kodak24, McMaster, and RNI6. FFDNet outperforms state-of-the-art denoisers such as BM3D, WNNM, MLP, TNRD, and DnCNN in terms of PSNR and visual quality. It is also effective in handling spatially variant noise by specifying a non-uniform noise level map. The model is flexible and efficient, making it suitable for practical denoising applications. FFDNet is trained with a noise level map and uses orthogonal initialization to avoid visual artifacts. It is compared with other methods in terms of noise level sensitivity and real noisy image denoising, showing robust performance across different noise levels and image structures. The results demonstrate that FFDNet is a highly effective and efficient solution for image denoising.