This paper shows that masked autoencoders (MAE) are scalable self-supervisedlearners for computer vision. Our MAE approach is simple: we mask randompatches of the input image and reconstruct the missing pixels. It is based ontwo core designs. First, we develop an asymmetric encoder-decoder architecture,with an encoder that operates only on the visible subset of patches (withoutmask tokens), along with a lightweight decoder that reconstructs the originalimage from the latent representation and mask tokens. Second, we find thatmasking a high proportion of the input image, e.g., 75%, yields a nontrivialand meaningful self-supervisory task. Coupling these two designs enables us totrain large models efficiently and effectively: we accelerate training (by 3xor more) and improve accuracy. Our scalable approach allows for learninghigh-capacity models that generalize well: e.g., a vanilla ViT-Huge modelachieves the best accuracy (87.8%) among methods that use only ImageNet-1Kdata. Transfer performance in downstream tasks outperforms supervisedpre-training and shows promising scaling behavior.