Research map / Label-efficient and robust learning / Recognition and long-tail
Stare at What You See: Masked Image Modeling Without Reconstruction
Read the paperCheck 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 |
|---|---|---|---|
| Remove reconstruction on masked tokens and align visible student features to intact teacher features | MaskAlign with ViT-B · ImageNet-1K val | fine-tuning top-1 accuracy (%): 85.0 -> 85.4 | Helped |
| Use MaskAlign instead of MILAN | ViT-B/16 · MS-COCO | mask AP: 45.5 -> 45.7 | Helped |
| Use MaskAlign instead of MILAN | ViT-B · ADE20K | mIoU (%): 52.7 -> 52.1 | Hurt |
| Add masked modeling to feature alignment, using attentive masking at 60–70% rather than 0% (ablation) | MaskAlign with ViT-B/16 · ImageNet-1K | fine-tuning top-1 accuracy (%): 84.9 -> 85.3 | Helped |
| Use random rather than attentive masking at 70% mask ratio (ablation) | MaskAlign with ViT-B/16 · ImageNet-1K | fine-tuning top-1 accuracy (%): 85.3 -> 85.1 | Hurt |
| Use an extremely high 90% mask ratio rather than the 60–70% peak (ablation) | MaskAlign with ViT-B/16 · ImageNet-1K | fine-tuning top-1 accuracy (%): 85.3 -> 84.4 | Hurt |
| Aggregate multiple teacher levels dynamically rather than align only to the final teacher output (ablation) | MaskAlign with ViT-B/16 · ImageNet-1K | fine-tuning top-1 accuracy (%): 84.9 -> 85.3 | Helped |
| Use Dynamic Alignment rather than layer-wise alignment at Top 5 (ablation) | MaskAlign with ViT-B/16 · ImageNet-1K | fine-tuning top-1 accuracy (%): 85.1 -> 85.3 | Helped |
| Use an MLP adaptor rather than a linear adaptor (ablation) | MaskAlign with ViT-B/16 · ImageNet-1K | fine-tuning top-1 accuracy (%): 85.3 -> 85.2 | Hurt |
| Use BatchNorm rather than LayerNorm (ablation) | MaskAlign with ViT-B/16 · ImageNet-1K | fine-tuning top-1 accuracy (%): 85.3 -> 84.7 | Hurt |
| Remove target normalization rather than use LayerNorm (ablation) | MaskAlign with ViT-B/16 · ImageNet-1K | fine-tuning top-1 accuracy (%): 85.3 -> 84.3 | Hurt |
| Add the [CLS] token to alignment (ablation) | MaskAlign with ViT-B/16 · ImageNet-1K | fine-tuning top-1 accuracy (%): 85.3 -> 85.4 | Helped |
Closest papers by meaning
- From Prototypes to General Distributions: An Efficient Curriculum for Masked Image Modeling (CVPR 2025 · similarity 0.89)
- Reconstruction Target Matters in Masked Image Modeling for Cross-Domain Few-Shot Learning (AAAI 2025 · similarity 0.88)
- Contrastive Tuning: A Little Help to Make Masked Autoencoders Forget (AAAI 2024 · similarity 0.88)
- The Dynamic Duo of Collaborative Masking and Target for Advanced Masked Autoencoder Learning (AAAI 2025 · similarity 0.88)
- Morphing Tokens Draw Strong Masked Image Models (ICLR 2025 · similarity 0.88)
- Learning Mask Invariant Mutual Information for Masked Image Modeling (ICLR 2025 · similarity 0.88)
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.