Architecture
aka ViT
Vision Transformer
Transformer that treats image patches as tokens — no convolutions required.
Definition
Vision Transformers split an image into 16×16 patches, linearly embed them and feed the sequence into a standard transformer. With enough pre-training data they outperform CNNs and have become the default backbone for multimodal and image-classification models.
Common use cases
- Image classification
- Multimodal backbones
- Detection