Research map / Architectures and efficiency / Attention and transformers
MixMAE: Mixed and Masked Autoencoder for Efficient Pretraining of Hierarchical Vision Transformers
Jihao Liu, Xin Huang, Jinliang Zheng, Yu Liu, Hongsheng Li
This paper tried a common change: Train longer
Train longer, from 300 to 600 epochs · ImageNet-1K · top-1 accuracy: 84.6 -> 85.1 Improved
What 387 accepted papers report for this change
- 211 improved
- 62 no clear effect
- 30 mixed
- 84 hurt
- Lightweight Federated Incremental Learning via Decoupled Replay ICML 2026 · increase local training epochs from 5 to 40 · CIFAR10 (α=1.0) · test accuracyImproved
- When Do Hallucinations Arise? A Graph Perspective on the Evolution of Path Reuse and Path Compression ICML 2026 · train longer on the full path search space (overfitting stage) · SBM graph, 10 communities, pin=0.6 · accuracy: 1.0 -> 0.98Hurt
- Causes and Consequences of Representational Similarity in Machine Learning Models ICML 2026 · train longer on LLM continued pretraining to test insufficient adaptation · perplexity across overlap settings · perplexityImproved
See this change in detailMatched by the wording of each paper's own result, read by AI from the full text. Papers report gains more often than failures, so this shows what gets published, not the chance a change works.
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Use mixed visible tokens and dual reconstruction instead of single-image masked reconstruction (ablation) | MixMAE with Swin-B/W14 · ImageNet-1K | top-1 accuracy: 84.0 -> 84.6 | Improved |
| Use mixed visible tokens and dual reconstruction instead of single-image masked reconstruction (ablation) | MixMAE with Swin-B/W14 · ADE20K | mIoU: 47.3 -> 49.9 | Improved |
| Fill masked areas with visible tokens from another image rather than zeros (ablation) | MixMAE · ImageNet-1K | top-1 accuracy: 84.1 -> 84.6 | Improved |
| Fill masked areas with visible tokens from another image rather than shuffled tokens (ablation) | MixMAE · ImageNet-1K | top-1 accuracy: 82.6 -> 84.6 | Improved |
| Fill masked areas with visible tokens from another image rather than zoomed-and-cropped image tokens (ablation) | MixMAE · ImageNet-1K | top-1 accuracy: 83.5 -> 84.6 | Improved |
| Use a shared learnable token rather than zeros for filling (ablation) | MixMAE · ADE20K | mIoU: 48.0 -> 48.9 | Improved |
| Mix 4 images (0.75 masking ratio) rather than 2 images (0.5) (ablation) | MixMAE · ImageNet-1K | top-1 accuracy: 84.6 -> 84.8 | Improved |
| Use masked self-attention to separate image groups (ablation) | MixMAE · ImageNet-1K | top-1 accuracy: 84.4 -> 85.1 | Improved |
| Train longer, from 300 to 600 epochs (ablation) | MixMAE · ImageNet-1K | top-1 accuracy: 84.6 -> 85.1 | Improved |
| Train longer, from 600 to 900 epochs (ablation) | MixMAE · ImageNet-1K | top-1 accuracy: 85.1 -> 85.1 | No clear effect |
| Train longer, from 300 to 900 epochs (ablation) | MixMAE · ADE20K | mIoU: 49.9 -> 51.0 | Improved |
| Replace mixed-image filling with [MASK] tokens for a 0.75 masking ratio (ablation) | MixMAE · ImageNet-1K | top-1 accuracy: 84.8 -> 84.4 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SimMIM — SimMIM: A Simple Framework for Masked Image Modeling (CVPR 2022) · better on 4 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Mixed Autoencoder for Self-Supervised Visual Representation Learning (CVPR 2023 · similarity 0.90)
- OmniMAE: Single Model Masked Pretraining on Images and Videos (CVPR 2023 · similarity 0.88)
- Morphing Tokens Draw Strong Masked Image Models (ICLR 2025 · similarity 0.86)
- Improved Baselines with Representation Autoencoders (NeurIPS 2026 · similarity 0.86)
- Closest Neighbors are Harmful for Lightweight Masked Auto-encoders (CVPR 2025 · similarity 0.86)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.86)
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.
