U²-Net is a novel deep network architecture designed for salient object detection (SOD). It features a two-level nested U-structure, enabling the network to capture rich local and global information without relying on pre-trained backbones from image classification tasks. The architecture uses ReSidual U-blocks (RSU) to extract multi-scale features while maintaining high resolution and reducing computational costs. U²-Net can be trained from scratch and achieves competitive performance on six SOD datasets. Two versions of the model are provided: a full-size version (176.3 MB, 30 FPS) and a smaller version (4.7 MB, 40 FPS). Both models demonstrate strong performance in terms of accuracy and efficiency. The network's design allows for flexibility in adapting to different environments with minimal performance loss. Experimental results show that U²-Net outperforms many state-of-the-art methods in terms of various evaluation metrics, including precision-recall curves, maxFβ, MAE, and boundary quality. Qualitative comparisons further highlight the model's ability to accurately detect salient objects in diverse scenarios, including small and large targets, targets touching image borders, and complex backgrounds. The paper concludes that U²-Net is a promising approach for SOD, offering a balance between performance and efficiency.U²-Net is a novel deep network architecture designed for salient object detection (SOD). It features a two-level nested U-structure, enabling the network to capture rich local and global information without relying on pre-trained backbones from image classification tasks. The architecture uses ReSidual U-blocks (RSU) to extract multi-scale features while maintaining high resolution and reducing computational costs. U²-Net can be trained from scratch and achieves competitive performance on six SOD datasets. Two versions of the model are provided: a full-size version (176.3 MB, 30 FPS) and a smaller version (4.7 MB, 40 FPS). Both models demonstrate strong performance in terms of accuracy and efficiency. The network's design allows for flexibility in adapting to different environments with minimal performance loss. Experimental results show that U²-Net outperforms many state-of-the-art methods in terms of various evaluation metrics, including precision-recall curves, maxFβ, MAE, and boundary quality. Qualitative comparisons further highlight the model's ability to accurately detect salient objects in diverse scenarios, including small and large targets, targets touching image borders, and complex backgrounds. The paper concludes that U²-Net is a promising approach for SOD, offering a balance between performance and efficiency.