Noise2Void is a novel training method for denoising convolutional neural networks (CNNs) that requires only a set of single, noisy images. Unlike traditional methods that need clean target images or noisy image pairs, Noise2Void (N2V) trains directly on the data to be denoised. This makes it applicable in scenarios where clean target images are not available, such as in biomedical imaging. N2V is based on the assumption that the signal in an image is not statistically independent and that noise is conditionally independent given the signal. It uses a blind-spot network architecture, where the receptive field of each pixel excludes the pixel itself, preventing the network from learning the identity mapping. This allows the network to learn to remove pixel-wise independent noise. N2V was evaluated on the BSD68 dataset and simulated microscopy data, and compared to traditional training, NOISE2NOISE (N2N) training, and non-trained methods like BM3D and non-local means. Results showed that N2V performs well, with denoising performance only moderately dropping compared to traditional methods. It was also applied to three biomedical datasets, including cryo-TEM images and fluorescence microscopy data, where traditional training was not feasible due to the lack of ground truth data. N2V demonstrated strong performance in these cases, outperforming BM3D. The method was implemented using a U-Net architecture with batch normalization and a linear activation function in the last layer. Training involved masking a random pixel in each input patch to create a blind-spot, allowing the network to learn denoising without relying on clean data. The results showed that N2V can effectively denoise images, even in the presence of structured noise, and outperformed traditional methods in many cases. However, it is not suitable for images with high irregularities that are difficult to predict. N2V provides a powerful alternative for training denoising networks, especially in scenarios where clean data is not available.Noise2Void is a novel training method for denoising convolutional neural networks (CNNs) that requires only a set of single, noisy images. Unlike traditional methods that need clean target images or noisy image pairs, Noise2Void (N2V) trains directly on the data to be denoised. This makes it applicable in scenarios where clean target images are not available, such as in biomedical imaging. N2V is based on the assumption that the signal in an image is not statistically independent and that noise is conditionally independent given the signal. It uses a blind-spot network architecture, where the receptive field of each pixel excludes the pixel itself, preventing the network from learning the identity mapping. This allows the network to learn to remove pixel-wise independent noise. N2V was evaluated on the BSD68 dataset and simulated microscopy data, and compared to traditional training, NOISE2NOISE (N2N) training, and non-trained methods like BM3D and non-local means. Results showed that N2V performs well, with denoising performance only moderately dropping compared to traditional methods. It was also applied to three biomedical datasets, including cryo-TEM images and fluorescence microscopy data, where traditional training was not feasible due to the lack of ground truth data. N2V demonstrated strong performance in these cases, outperforming BM3D. The method was implemented using a U-Net architecture with batch normalization and a linear activation function in the last layer. Training involved masking a random pixel in each input patch to create a blind-spot, allowing the network to learn denoising without relying on clean data. The results showed that N2V can effectively denoise images, even in the presence of structured noise, and outperformed traditional methods in many cases. However, it is not suitable for images with high irregularities that are difficult to predict. N2V provides a powerful alternative for training denoising networks, especially in scenarios where clean data is not available.