This paper introduces BiRefNet, a novel bilateral reference framework for high-resolution dichotomous image segmentation (DIS). The framework consists of a localization module (LM) and a reconstruction module (RM), with the RM utilizing bilateral reference (BiRef) for reconstruction. BiRef combines hierarchical image patches as source references and gradient maps as target references. Auxiliary gradient supervision is introduced to enhance focus on detailed regions. Practical training strategies are also proposed to improve map quality and training efficiency. BiRefNet outperforms state-of-the-art methods on DIS, HRSOD, and COD benchmarks, achieving significant improvements in metrics such as S-measure, F-measure, and HCE. The framework is evaluated on four tasks, demonstrating its effectiveness in high-resolution segmentation. BiRefNet also shows strong generalization capabilities and is applied to various practical scenarios, including crack detection and object extraction. The model is efficient, with high performance and low computational cost. Third-party applications based on BiRefNet have been developed, highlighting its potential for real-world use. The proposed method provides a simple yet effective solution for high-resolution image segmentation, with promising results across multiple tasks and applications.This paper introduces BiRefNet, a novel bilateral reference framework for high-resolution dichotomous image segmentation (DIS). The framework consists of a localization module (LM) and a reconstruction module (RM), with the RM utilizing bilateral reference (BiRef) for reconstruction. BiRef combines hierarchical image patches as source references and gradient maps as target references. Auxiliary gradient supervision is introduced to enhance focus on detailed regions. Practical training strategies are also proposed to improve map quality and training efficiency. BiRefNet outperforms state-of-the-art methods on DIS, HRSOD, and COD benchmarks, achieving significant improvements in metrics such as S-measure, F-measure, and HCE. The framework is evaluated on four tasks, demonstrating its effectiveness in high-resolution segmentation. BiRefNet also shows strong generalization capabilities and is applied to various practical scenarios, including crack detection and object extraction. The model is efficient, with high performance and low computational cost. Third-party applications based on BiRefNet have been developed, highlighting its potential for real-world use. The proposed method provides a simple yet effective solution for high-resolution image segmentation, with promising results across multiple tasks and applications.