Alexey Dosovitskiy et al. at Google Research published An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (ViT), showing that slicing images into 16x16 patch tokens and feeding them into a standard Transformer outperforms ResNets when pretrained on large image corpora.
Vision Transformer (ViT)
Also known as: ViT · MAE · Masked Autoencoder · NaViT
Plain English
An architecture that applies the Transformer directly to images by slicing a picture into a grid of small square patches and treating each patch like a word in a sentence.
Technical Overview
Proposed by Dosovitskiy et al., ViT flattens P x P image patches into linear token projections with positional embeddings and processes them with standard Transformer encoder blocks. Lacking the built-in translation equivariance of CNNs, ViTs scale strongly with large pretraining datasets or masked patch reconstruction (MAE) and serve as the universal visual backbone for CLIP, DINO, SAM, and VLMs.
Timeline (4)
2020
Paper2021
PaperKaiming He et al. at Meta AI introduced Masked Autoencoders (MAE), masking 75% of input image patches and applying the ViT encoder only to the visible 25% before a lightweight decoder reconstructs the missing pixels—cutting self-supervised pretraining compute by over 3x.
2023
PaperDehghani et al. at Google DeepMind published NaViT (Patch n' Pack), packing variable-resolution and native-aspect-ratio patch sequences into single training batches with factorized 2D positional embeddings rather than distorting images into fixed squares.
2025
PaperGoogle DeepMind's SigLIP 2 unified NaViT aspect-preserving patch encoders with sigmoid contrastive, captioning, and self-distillation objectives, providing open-weight ViT backbones tailored for dense OCR, localization, and vision-language models.
See also (10)
Open full graphVision Transformer (ViT)