Research map / Architectures and efficiency / Pruning and efficient tuning
TinyMIM: An Empirical Study of Distilling MIM Pre-Trained Models
Sucheng Ren, Fangyun Wei, Zheng Zhang, Han Hu
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| Distill token relations rather than class tokens or features (ablation) | TinyMIM ViT-T, ViT-S, and ViT-B · ImageNet-1K | top-1 accuracy: 75.8 (+4.2); 83.0 (+3.1); 85.0 (+1.6) | Improved |
| Distill token relations rather than class tokens or features (ablation) | TinyMIM ViT-B · ADE20K | mIoU: 52.2 (+4.1) | Improved |
| Add MAE reconstruction loss to class-token distillation (ablation) | TinyMIM ViT-B · ImageNet-1K | top-1 accuracy: 80.6 -> 82.1 | Improved |
| Distill FFN or attention features before, rather than after, the residual connection (ablation) | TinyMIM ViT-B · ImageNet-1K | top-1 accuracy: 84.2 -> 81.8; 84.1 -> 81.3 | Hurt |
| Apply softmax to the Q-K and V-V relation targets (ablation) | TinyMIM ViT-B · ImageNet-1K | top-1 accuracy: 84.4 -> 84.6 | Improved |
| Use Q-K and V-V relations instead of Q-Q, K-K, and V-V relations (ablation) | TinyMIM ViT-B · ImageNet-1K | top-1 accuracy: 84.4 -> 84.4 | No clear effect |
| Select teacher block 18 rather than block 24 (ablation) | ViT-B distilled from MAE ViT-L · ImageNet-1K | top-1 accuracy: 84.4 -> 84.6 | Improved |
| Select teacher block 18 rather than block 24 (ablation) | ViT-B distilled from MAE ViT-L · ADE20K | mIoU: 50.0 -> 52.2 | Improved |
| Use MAE ViT-L directly instead of MAE ViT-B as teacher (ablation) | ViT-S · ImageNet-1K | top-1 accuracy: 82.3 -> 82.1 | Hurt |
| Use sequential teacher ViT-L -> TinyMIM ViT-B instead of MAE ViT-L directly (ablation) | ViT-S · ImageNet-1K | top-1 accuracy: 82.1 -> 82.6 | Improved |
| Use sequential teacher MAE ViT-B -> TinyMIM ViT-S instead of MAE ViT-B directly (ablation) | ViT-T · ImageNet-1K | top-1 accuracy: 74.4 -> 75.0 | Improved |
| Use a raw image rather than a masked image as distillation input (ablation) | TinyMIM ViT-B · ImageNet-1K | top-1 accuracy: 83.9 -> 84.6 | Improved |
Closest papers by meaning
- TinyViT: Fast Pretraining Distillation for Small Vision Transformers (ECCV 2022 · similarity 0.91)
- On Data Scaling in Masked Image Modeling (CVPR 2023 · similarity 0.90)
- Asymmetric Masked Distillation for Pre-Training Small Foundation Models (CVPR 2024 · similarity 0.90)
- Closest Neighbors are Harmful for Lightweight Masked Auto-encoders (CVPR 2025 · similarity 0.89)
- Morphing Tokens Draw Strong Masked Image Models (ICLR 2025 · similarity 0.89)
- From Prototypes to General Distributions: An Efficient Curriculum for Masked Image Modeling (CVPR 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.
