Research map / Architectures and efficiency / Attention and transformers
Masked Autoencoders Are Scalable Vision Learners
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| encode only visible patches and add mask tokens in the lightweight decoder (ablation) | MAE ViT-L · ImageNet-1K pre-training | training speedup: 2.8× | Helped |
| put mask tokens in the encoder (ablation) | MAE ViT-L/16 · ImageNet-1K | linear probing accuracy: 73.5 -> 59.6 | Hurt |
| normalize pixel reconstruction targets per patch (ablation) | MAE ViT-L/16 · ImageNet-1K | fine-tuning accuracy: 84.9 -> 85.4 | Helped |
| use dVAE tokens instead of unnormalized pixels as targets (ablation) | MAE ViT-L/16 · ImageNet-1K | linear probing accuracy: 73.5 -> 71.6 | Hurt |
| use dVAE tokens instead of normalized pixels as targets (ablation) | MAE · ImageNet-1K, COCO, and ADE20K | transfer performance: 0.0, -0.2, 0.0, 0.0, -0.1, 0.0, -0.2 | No clear effect |
| use PCA coefficients as reconstruction targets (ablation) | MAE ViT-L/16 · ImageNet-1K | fine-tuning accuracy: 84.9 -> 84.6 | Hurt |
| add color jitter to cropping augmentation (ablation) | MAE ViT-L/16 · ImageNet-1K | fine-tuning accuracy: 84.9 -> 84.3 | Hurt |
| remove data augmentation (ablation) | MAE ViT-L/16 · ImageNet-1K | linear probing accuracy: 73.5 -> 65.7 | Hurt |
| use block-wise masking at ratio 75 instead of random masking (ablation) | MAE ViT-L/16 · ImageNet-1K | fine-tuning accuracy: 84.9 -> 82.8 | Hurt |
| use grid-wise masking instead of random masking (ablation) | MAE ViT-L/16 · ImageNet-1K | linear probing accuracy: 73.5 -> 66.0 | Hurt |
| reduce decoder depth from 8 blocks to 1 (ablation) | MAE ViT-L/16 · ImageNet-1K | linear probing accuracy: 73.5 -> 65.5 | Hurt |
| reduce decoder depth from 8 blocks to 1 (ablation) | MAE ViT-L/16 · ImageNet-1K pre-training | training speedup: 2.8× -> 3.7× | Helped |
Closest papers by meaning
- Self-Guided Masked Autoencoder (NeurIPS 2024 · similarity 0.93)
- The Dynamic Duo of Collaborative Masking and Target for Advanced Masked Autoencoder Learning (AAAI 2025 · similarity 0.92)
- Learning Mask Invariant Mutual Information for Masked Image Modeling (ICLR 2025 · similarity 0.92)
- Extending Video Masked Autoencoders to 128 frames (NeurIPS 2024 · similarity 0.92)
- Closest Neighbors are Harmful for Lightweight Masked Auto-encoders (CVPR 2025 · similarity 0.92)
- Contrastive Tuning: A Little Help to Make Masked Autoencoders Forget (AAAI 2024 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.