BiRefNet (Bilateral Reference Network) is the official implementation of "Bilateral Reference for High-Resolution Dichotomous Image Segmentation" (CAAI AIR 2024), developed by researchers from Nankai University, Northwestern Polytechnical University, and collaborators including Shanghai AI Laboratory. It targets high-resolution dichotomous image segmentation (DIS) by combining a localization module with a bilateral reference module that fuses low-level detail cues with high-level semantic context, letting it produce precise object masks even on very large images. The project ships pretrained weights for general-purpose segmentation, matting, high-resolution (2K) inference, salient object detection, and camouflaged object detection, and can be loaded in a single line via Hugging Face Transformers (AutoModelForImageSegmentation). It is actively maintained, with ONNX export, Colab demos, a Hugging Face Space, and a fal.ai inference API, and has been adopted by third-party tools for background removal and matting.