Masked Autoencoders Are Scalable Vision Learners

Vision self-supervised learners had not scaled comparably to NLP because masking was applied in a way that wasted encoder compute.

Mask random patches and reconstruct missing pixels, with two design choices: an ASYMMETRIC encoder-decoder where the encoder sees only visible patches (no mask tokens) and a lightweight decoder reconstructs from latents plus mask tokens; and HIGH masking ratio (75%), which makes the task non-trivial. Because the encoder never processes masked tokens, encoder FLOPs scale with the visible fraction.

Accelerates training by 3x or more and improves accuracy versus prior self-supervised approaches; a vanilla ViT-Huge reaches 87.8% on ImageNet-1K using only ImageNet-1K data. These are TRAINING numbers - there is no inference speedup in the paper, so nothing here transfers to a serving decision.

Venue CVPR 2022 (crossref container '2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition'). MAE is an ENCODER paper, not a VLM paper: its relevance to this repo is the asymmetric visible-subset encoder, which is the same shape of idea as visual token sparsification (pap records dynamicvit, visionzip, sparsevlm) - compute proportional to kept tokens rather than all tokens. That is the general principle behind most vision-token reduction, so MAE belongs in the lineage even though no VLM ships it as a projector. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'main'. This is STRUCTURAL rather than researched: cvpr has no Findings volume and no workshop track in which a paper could appear, so 'main' is the only possible honest value and no per-paper evidence was needed. The field is annotated here so that 'findings' and 'workshop' remain MEANINGFUL - a filter for findings=true must return only the papers that actually are findings, which requires every other paper to be explicitly NOT findings.