SAM-Adapter adapts Meta AI’s Segment Anything foundation model to underperformed downstream scenes by integrating adapters and prompts into the model’s architecture. The approach enables fine-tuning SAM on specialized tasks such as camouflage detection, shadow detection, polyp segmentation, and other medical imaging applications. The project supports multiple SAM versions (SAM, SAM2, SAM3) and provides configurations for ViT-H, ViT-L, and ViT-B backbones.