Skip to content

Paper 079 / Google Research

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

Vision Transformer applies a largely standard Transformer encoder to sequences of image patches and studies when large-scale pretraining makes it competitive with convolutional networks.

UPDATED 16 Jul 2026SOURCE-LED REVIEWRESEARCH REVIEW
01

Paper, researchers, and primary source

Major lab research / computer_vision

Vision Transformer applies a largely standard Transformer encoder to sequences of image patches and studies when large-scale pretraining makes it competitive with convolutional networks.

CONTRIBUTION / 01

Contribution 1

Represented images as sequences of embedded fixed-size patches.

CONTRIBUTION / 02

Contribution 2

Applied Transformer encoders without convolutional feature hierarchies.

CONTRIBUTION / 03

Contribution 3

Showed strong transfer after pretraining on sufficiently large image datasets.

02

Research context

computer_vision / 2020

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale places vision transformer inside the broader computer vision discussion at Google Research, with image patches supplying a second analytical lens. The editorial sequence connects three claims: Represented images as sequences of embedded fixed-size patches; Applied Transformer encoders without convolutional feature hierarchies; and Showed strong transfer after pretraining on sufficiently large image datasets. The combination matters because transformer only has meaning under the paper's stated setup. Operationally, the record points to one consequence: the reported advantage depends strongly on pretraining scale and resolution, while patching can miss useful inductive bias for smaller-data settings.

03

Methods and evidence reading

1 cataloged method notes

METHOD / 01

Method 1

The experimental design in An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale tests represented images as sequences of embedded fixed-size patches and applied transformer encoders without convolutional feature hierarchies against the paper's documented baselines, datasets, model variants, or systems workloads.

How to read the evidence

For An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, the useful question is what the documented setup supports and where transfer remains untested. Its evidence chain begins with two documented moves: Represented images as sequences of embedded fixed-size patches; and Applied Transformer encoders without convolutional feature hierarchies. Reported evidence then addresses: Showed strong transfer after pretraining on sufficiently large image datasets. The resulting interpretation is practical but conditional: the reported advantage depends strongly on pretraining scale and resolution, while patching can miss useful inductive bias for smaller-data settings. Its boundary is that the vision transformer comparison in An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale is interpretable only alongside reported datasets, selected metrics, prompt protocol, domain coverage, image resolution, and input modalities, which limits claims about unseen deployments. Any extension should report how altered image patches conditions affect the original vision transformer result. For a follow-on study of An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, pair vision transformer measurements with image patches slices and preserve negative examples around Applied Transformer encoders without convolutional feature hierarchies as first-class evidence.

04

Findings in the source record

1 paper-specific findings

  1. The reported evidence in An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale supports showed strong transfer after pretraining on sufficiently large image datasets.
05

Practical implication for AI builders

Google Research / 2020

06

Proposed BrokenGPT application

Research blueprint / proposed status

INTEGRATION POINT / 01

Proposed route placement / vision transformer

For a proposed BrokenGPT experiment based on An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, compare ViT and convolutional encoders on the same licensed image corpus, resolution, augmentation, latency, and transfer tasks before choosing a visual backbone. Keep the vision transformer path isolated, versioned, and attributable to this research record.

VALIDATION METRIC / 02

Proposed acceptance test / image patches

Validate the proposed vision transformer route against the paper's reported outcome: Showed strong transfer after pretraining on sufficiently large image datasets. Measure transfer robustness, boundary errors, task accuracy, and calibration for the An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale candidate, then isolate image patches regressions before judging the proposed vision transformer route.

TRADEOFF / 03

Proposed decision boundary / transformer

Balance resolution, latency, and domain shift before promoting the proposed transformer design. Because for a later vision transformer implementation, source provenance, memorization, media rights, representation bias, accessibility, and misuse define unresolved boundaries that require direct observation, adoption remains conditional on replay under BrokenGPT's selected model, runtime, and policy configuration.

07

Limitations, verification, and source

Boundaries recorded with the paper

Limitations

  • The vision transformer comparison in An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale is interpretable only alongside reported datasets, selected metrics, prompt protocol, domain coverage, image resolution, and input modalities, which limits claims about unseen deployments.
  • For a later vision transformer implementation, source provenance, memorization, media rights, representation bias, accessibility, and misuse define unresolved boundaries that require direct observation.

PRIMARY SOURCES

  1. 01
    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    Google Research — Primary primary arXiv paper / 22 October 2020 / Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, and 9 more

Related research reviews

View all 100 credited research papers

STRAIGHT ANSWERS

Frequently asked questions

01What does An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale study?

Vision Transformer applies a largely standard Transformer encoder to sequences of image patches and studies when large-scale pretraining makes it competitive with convolutional networks.

02Which methods does An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale use?

The experimental design in An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale tests represented images as sequences of embedded fixed-size patches and applied transformer encoders without convolutional feature hierarchies against the paper's documented baselines, datasets, model variants, or systems workloads.

03What does An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale report?

The reported evidence in An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale supports showed strong transfer after pretraining on sufficiently large image datasets.

04What is the proposed BrokenGPT application for An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale?

Proposed: compare ViT and convolutional encoders on the same licensed image corpus, resolution, augmentation, latency, and transfer tasks before choosing a visual backbone.

MAJOR LAB RESEARCH / PAPER 079

Continue after An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

After An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, browse the full index for adjacent computer_vision research and work from Google Research.

Open the paper index