Modern ViT-Adapter

Built on the ViT-Adapter architecture, this project modernizes the original codebase for PyTorch 2.x+ with easy installation.

The versions mmcv==1.4.2, mmdet==2.22.0, and mmseg==0.20.2 are kept, with compatibility updates for the latest PyTorch. See the log of these fixes here.

News🔥

2026-05-03 Added support for ViT-CoMer and ViT-Split.

2026-04-30 Added support for MaskDINO and two adapters for DINOv3. See Release v1.1.0.

Installation

See INSTALL.md.

Train and test

Important: To activate the environment, you must perform both of the following steps every time:

1. Use your environment manager to activate the environment (for example: micromamba activate torch29).
2. Run source env.sh to link to mmcv, mmdet and mmseg.

For training and testing, the usage is exactly the same with the original ViT-Adapter project, see detection/README.md and segmentation/README.md. Apart from the original usage, we also support the DINOv3 backbones. Please refer to scripts/example.train.sh and scripts/example.test.sh.

Todo

• Support ViT-CoMer - another excellent adapter designed for ViT-based backbones
• Support MaskDINO - a solid task head for detection and segmentation.