czczup/ViT-Adapter

Health v2 · maintenance only

Flags: no_license

Adoption not part of the score

1503 stars · 155 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

Official PyTorch implementation of ViT-Adapter, an ICLR 2023 Spotlight paper introducing a pre-training-free adapter that lets plain Vision Transformers achieve strong dense prediction performance. It provides models and configs for object detection, instance segmentation, and semantic segmentation built on MMDetection and MMSegmentation.

Use cases

• run object detection with a plain ViT backbone on COCO
• perform semantic segmentation on ADE20K or Cityscapes with ViT-Adapter
• reproduce ICLR 2023 paper results for dense predictions
• fine-tune a plain ViT for instance and panoptic segmentation
• compare adapter-based ViT against vision-specific transformers
• use pretrained ViT-Adapter checkpoints in detection pipelines

When to choose

• you need state-of-the-art dense prediction results with a plain ViT backbone
• you already use MMDetection or MMSegmentation ecosystems
• you want pretrained checkpoints for detection and segmentation research

When to avoid

• you need a lightweight inference-only library without the OpenMMLab stack
• your project requires a permissive, clearly defined license
• you work outside PyTorch or lack GPU resources

Facets

library · maturity stable

machine-learning computer-vision image-processing computer-vision deep-learning machine-learning python vision-transformer adapter object-detection semantic-segmentation instance-segmentation panoptic-segmentation iclr-2023 mmdetection dense-prediction gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="czczup/ViT-Adapter")