Research map / Architectures and efficiency / Attention and transformers
MetaFormer Baselines for Vision
Weihao Yu, Chenyang Si, Pan Zhou, Mi Luo, Yichen Zhou, Jiashi Feng, Shuicheng Yan, Xinchao Wang
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 |
|---|---|---|---|
| use random mixing in the top two stages instead of identity mapping | IdentityFormer-S12 · ImageNet-1K val | top-1 accuracy: 74.6 -> 76.6 | Improved |
| scale identity-mapping MetaFormer up to M48 | IdentityFormer · ImageNet-1K val | top-1 accuracy: 74.6 -> 80.4 | Improved |
| use pooling instead of random mixing | RandFormer-M48 · ImageNet-1K val | top-1 accuracy: 81.4 -> 82.6 | Improved |
| add a 4-layer conv stem (ablation) | isotropic IdentityFormer-S · ImageNet-1K val | top-1 accuracy: 68.2 -> 75.4 | Improved |
| replace conv token mixer with attention in top two stages (CAFormer vs ConvFormer) | ConvFormer-S18 · ImageNet-1K val | top-1 accuracy: 83.0 -> 83.6 | Improved |
| add ImageNet-21K pretraining | CAFormer-B36 · ImageNet-1K val | top-1 accuracy: 85.5 -> 87.4 | Improved |
| use StarReLU instead of ReLU (ablation) | ConvFormer-S18 · ImageNet-1K val | top-1 accuracy: 82.1 -> 83.0 | Improved |
| use StarReLU instead of GELU (ablation) | ConvFormer-S18 · ImageNet-1K val | top-1 accuracy: 82.7 -> 83.0 | Improved |
| use StarReLU instead of GELU (ablation) | CAFormer-S18 · ImageNet-1K val | top-1 accuracy: 83.4 -> 83.6 | Improved |
| use StarReLU instead of Squared ReLU (ablation) | ConvFormer-S18 · ImageNet-1K val | top-1 accuracy: 82.6 -> 83.0 | Improved |
| use ResScale instead of LayerScale (ablation) | CAFormer-S18 · ImageNet-1K val | top-1 accuracy: 83.0 -> 83.6 | Improved |
| disable biases in blocks (ablation) | ConvFormer-S18 · ImageNet-1K val | top-1 accuracy: 83.0 -> 83.0 | No clear effect |
Closest papers by meaning
- FFT-Based Dynamic Token Mixer for Vision (AAAI 2024 · similarity 0.84)
- One Meta-tuned Transformer is What You Need for Few-shot Learning (ICML 2024 · similarity 0.83)
- LMUFormer: Low Complexity Yet Powerful Spiking Model With Legendre Memory Units (ICLR 2024 · similarity 0.82)
- TokenFormer: Rethinking Transformer Scaling with Tokenized Model Parameters (ICLR 2025 · similarity 0.82)
- Hydra: Bidirectional State Space Models Through Generalized Matrix Mixers (NeurIPS 2024 · similarity 0.82)
- Morphing Tokens Draw Strong Masked Image Models (ICLR 2025 · similarity 0.82)
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.
