BiRefNet

BiRefNet is a dichotomous image segmentation model designed for high-resolution input, introduced in the paper "Bilateral Reference for High-Resolution Dichotomous Image Segmentation" (CAAI AIR 2024, DOI 10.26599/AIR.2024.9150038). It achieves state-of-the-art performance on dichotomous image segmentation (DIS), high-resolution salient object detection (HRSOD), and camouflaged object detection (COD).

Architecture and Capabilities

BiRefNet employs a bilateral reference framework comprising a localization module (LM) for object localization using global semantic information and a reconstruction module (RM) that leverages bilateral references: hierarchical image patches as source reference and gradient maps as target reference. Auxiliary gradient supervision is applied to improve fine-detail regions. The model was trained on the DIS-TR dataset and validated on DIS-TEs and DIS-VD.

Notable Performance and Optimizations

• As of June 2025, the refine_foreground operation was accelerated by approximately 8× to ~80 ms on an NVIDIA 5090 GPU using a GPU implementation of fast-fg-est.
• The Swin Transformer attention module was upgraded to PyTorch’s official SDPA (September 2025), reducing memory cost and offering potential speedups during training and inference.
• The model has been updated to version v7 (arXiv 2401.03407, July 2025) with a corrected figure layout.

Visual Examples

DIS Sample 1	DIS Sample 2

The image below shows the online inference GUI demo on Hugging Face, which supports adjustable resolutions.

Additional Resources

Source code, weights, and a model zoo are available on the GitHub repository. The full paper provides further details on the bilateral reference design and experimental benchmarks.