The paper introduces U²-Net, a novel deep network architecture designed for salient object detection (SOD). The architecture is characterized by a two-level nested U-structure, which allows the network to capture rich multi-scale contextual information from different scales. This design enables the network to achieve competitive performance without using pre-trained backbones from image classification tasks. The authors propose a new residual U-block (RSU) that enhances the extraction of multi-scale features while maintaining high resolution. Two versions of U²-Net are presented: a full-size model (176.3 MB, 30 FPS) and a smaller version (4.7 MB, 40 FPS). Both models demonstrate superior performance on six public SOD datasets compared to 20 state-of-the-art methods. The paper also includes ablation studies to validate the effectiveness of the proposed blocks, architecture, and backbone-free design. Experimental results show that U²-Net outperforms existing methods in terms of precision, recall, mean absolute error, weighted F-measure, structure measure, and relaxed boundary F-measure.The paper introduces U²-Net, a novel deep network architecture designed for salient object detection (SOD). The architecture is characterized by a two-level nested U-structure, which allows the network to capture rich multi-scale contextual information from different scales. This design enables the network to achieve competitive performance without using pre-trained backbones from image classification tasks. The authors propose a new residual U-block (RSU) that enhances the extraction of multi-scale features while maintaining high resolution. Two versions of U²-Net are presented: a full-size model (176.3 MB, 30 FPS) and a smaller version (4.7 MB, 40 FPS). Both models demonstrate superior performance on six public SOD datasets compared to 20 state-of-the-art methods. The paper also includes ablation studies to validate the effectiveness of the proposed blocks, architecture, and backbone-free design. Experimental results show that U²-Net outperforms existing methods in terms of precision, recall, mean absolute error, weighted F-measure, structure measure, and relaxed boundary F-measure.