Research map / Architectures and efficiency / Attention and transformers
Architecture-Agnostic Masked Image Modeling -- From ViT back to CNN
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 |
|---|---|---|---|
| Replace masked input pixels with RGB mean (ablation) | SimMIM-style ResNet-50 · ImageNet-100 | top-1 accuracy: 87.75 -> 88.19 | Helped |
| Add Fourier loss without adaptive frequency weighting (ablation) | A²MIM ResNet-50 · ImageNet-100 | top-1 accuracy: 88.19 -> 88.47 | Helped |
| Use adaptive weighting in the Fourier loss (ablation) | A²MIM ResNet-50 · ImageNet-100 | top-1 accuracy: 88.47 -> 88.73 | Helped |
| Add an intermediate mask token (ablation) | A²MIM ResNet-50 · ImageNet-100 | top-1 accuracy: 88.73 -> 88.86 | Helped |
| Add an intermediate mask token (ablation) | A²MIM ViT-B · ImageNet-1K | top-1 accuracy: 83.4 -> 83.5 | Helped |
| Replace the linear decoder with a 2-layer MLP (ablation) | A²MIM · ImageNet-1K | fine-tuning top-1 accuracy: 78.8 -> 78.8; 82.4 -> 82.4 | No clear effect |
| Use a 2-layer Transformer decoder instead of the linear decoder (ablation) | A²MIM · ImageNet-1K | fine-tuning top-1 accuracy: 78.8 -> 78.6; 82.4 -> 82.3 | Hurt |
| Add depth-wise convolution to the 2-layer MLP decoder (ablation) | A²MIM · ImageNet-1K | fine-tuning top-1 accuracy: 78.8 -> 78.9; 82.4 -> 82.5 | Helped |
| Add depth-wise convolution to the 2-layer Transformer decoder (ablation) | A²MIM · ImageNet-1K | fine-tuning top-1 accuracy: 78.6 -> 78.8; 82.3 -> 82.6 | Helped |
| Replace RGB target with HoG features (ablation) | A²MIM · ImageNet-1K | fine-tuning top-1 accuracy: 78.8 -> 78.9; 82.4 -> 82.6 | Helped |
| Replace RGB target with DINO features (ablation) | A²MIM · ImageNet-1K | fine-tuning top-1 accuracy: 78.8 -> 79.0; 82.4 -> 82.7 | Helped |
| Vary masked patch size (ablation) | A²MIM ResNet-50 · ImageNet-1K | top-1 accuracy: 78.2~/~78.6~/~78.8~/~78.7 | Mixed |
Closest papers by meaning
- Revealing the Dark Secrets of Masked Image Modeling (CVPR 2023 · similarity 0.90)
- Masked Image Modeling With Local Multi-Scale Reconstruction (CVPR 2023 · similarity 0.90)
- From Prototypes to General Distributions: An Efficient Curriculum for Masked Image Modeling (CVPR 2025 · similarity 0.90)
- Wavelet-Driven Masked Image Modeling: A Path to Efficient Visual Representation (AAAI 2025 · similarity 0.88)
- Pre-training with Random Orthogonal Projection Image Modeling (ICLR 2024 · similarity 0.88)
- Morphing Tokens Draw Strong Masked Image Models (ICLR 2025 · similarity 0.87)
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.