Masked Autoencoders Are Scalable Vision Learners
Summary
This paper shows that masked autoencoders (MAE) are scalable self-supervised learners for computer vision. The approach masks random patches of the input image and reconstructs the missing pixels, using an encoder that operates only on visible patches and a lightweight decoder. The authors find that masking a high proportion of the image, e.g., 75%, yields a nontrivial self-supervisory task, and they report that a vanilla ViT-Huge model achieves 87.8% accuracy among methods that use only ImageNet-1K data. [S1]
Claims
- id: C1 claim: 'MAE masks random image patches and reconstructs the missing pixels.' source: S1 evidence: "we mask random patches of the input image and reconstruct the missing pixels"
- id: C2 claim: 'The authors find that masking a high proportion of the image, e.g. 75%, is a strong self-supervisory task.' source: S1 evidence: "masking a high proportion of the input image, e.g., 75%"
- id: C3 claim: 'They report that a vanilla ViT-Huge model achieves 87.8% accuracy among ImageNet-1K-only methods.' source: S1 evidence: "a vanilla ViT-Huge model achieves the best accuracy (87.8%)"
Verification
- Code available: No official code URL was named on the abs page or HTML.
- Reproduced: not reproduced (2026-11-12). Code not run.
- Related work: Live note 2103-00020-clip.
- Limits: The 87.8% ImageNet-1K result for ViT-Huge and the 3x-or-more training speedup are author-reported (v3, 19 Dec 2021).
- License: CC BY 4.0 on the abs page (2026-11-12; icon by-4.0.png; http://creativecommons.org/licenses/by/4.0/). Submitted 11 Nov 2021 (v1); last revised 19 Dec 2021 (v3). Sources: [S1] https://arxiv.org/abs/2111.06377, v3 (19 Dec 2021). checked_on: 2026-11-12.
When to apply
Use when asking how masked autoencoding was scaled for vision Transformers, including 75% masking and an asymmetric encoder-decoder.
Facts
- id: F1 claim: 'MAE is an asymmetric masked autoencoder: the encoder sees only visible patches, and a lightweight decoder reconstructs missing pixels.' confidence: medium date: 2026-11-12 sources: [S1] license: CC BY 4.0
- id: F2 claim: 'The authors use a high masking ratio, for example 75%.' confidence: medium date: 2026-11-12 sources: [S1] license: CC BY 4.0
- id: F3 claim: 'They report 87.8% accuracy for a vanilla ViT-Huge trained with only ImageNet-1K data.' confidence: medium date: 2026-11-12 sources: [S1] license: CC BY 4.0
Sources
- [S1] arXiv 2111.06377, https://arxiv.org/abs/2111.06377. License: CC BY 4.0. v3 (19 Dec 2021). Accessed 2026-11-12.
License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted). Paper is CC BY 4.0.
License: CC-BY-4.0 · Summary by iinventedlight (AI-assisted). Source: Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick, Masked Autoencoders Are Scalable Vision Learners (CC BY 4.0).
Cite this
'iinventedlight (AI-assisted)'. 2026. "Masked Autoencoders Are Scalable Vision Learners." iinventedlight Library, iil:note:t1yag08y, version of 2026-11-12. https://iinventedlight.com/library/notes/2111-06377-mae/. License: CC-BY-4.0.
@misc{iil_note_t1yag08y,
author = {{'iinventedlight (AI-assisted)'}},
title = {{Masked Autoencoders Are Scalable Vision Learners}},
howpublished = {iinventedlight Library},
year = {2026},
date = {2026-11-12},
url = {https://iinventedlight.com/library/notes/2111-06377-mae/},
note = {Permanent ID iil:note:t1yag08y. License: CC-BY-4.0.},
}
Spotted a mistake? Use the MCP feedback tool or email hello@iinventedlight.com.
For language models and agents: start at /llms.txt | /agents.md | /.well-known/agent-library.json. This page in Markdown: /library/notes/2111-06377-mae.md.