Research map / Architectures and efficiency / Attention and transformers
HIRI-ViT: Scaling Vision Transformer With High Resolution Inputs
Ting Yao, Yehao Li, Yingwei Pan, Tao Mei
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 CFFN with depth-wise convolution instead of FFN (ablation) | M-ViT · ImageNet-1K val | top-1: 82.7 -> 82.9 | Improved |
| remove MHA in first two stages (ablation) | M-ViT · ImageNet-1K val | top-1: 82.9 -> 82.9 | No clear effect |
| replace LN with BN in first two stages (ablation) | M-ViT · ImageNet-1K val | top-1: 82.9 -> 83.0 | Improved |
| replace ViT-stem with Conv-stem (ablation) | M-ViT · ImageNet-1K val | top-1: 83.0 -> 83.4 | Improved |
| add inverted residual downsampling (ablation) | M-ViT · ImageNet-1K val | top-1: 83.4 -> 83.6 | Improved |
| increase input resolution 224 to 448 in four stages (ablation) | M-ViT · ImageNet-1K val | top-1 and GFLOPs: 83.6 -> 84.6; GFLOPs 4.7 -> 21.5 | Mixed |
| use five stages instead of four at 448 input (ablation) | M-ViT · ImageNet-1K val | top-1 and GFLOPs: 84.6 -> 84.0; GFLOPs 21.5 -> 5.8 | Mixed |
| replace Conv-stem with HR-stem (ablation) | HIRI-ViT-S · ImageNet-1K val | GFLOPs: 5.8 -> 5.1 | Improved |
| replace CFFN with two HR blocks in first two stages (ablation) | HIRI-ViT-S · ImageNet-1K val | top-1: 84.0 -> 84.1 | Improved |
| add EMA distillation (ablation) | HIRI-ViT-S · ImageNet-1K val | top-1: 84.1 -> 84.3 | Improved |
| use three Transformer stages instead of one at the end (ablation) | HIRI-ViT-S · ImageNet-1K val | top-1: 83.0 -> 84.3 | Improved |
| apply HIRI-ViT scaling to DaViT-Tiny at 448 (ablation) | DaViT-Tiny · ImageNet-1K val | top-1: 82.8 -> 83.3 | Improved |
Closest papers by meaning
- CvT: Introducing Convolutions to Vision Transformers (ICCV 2021 · similarity 0.91)
- FasterViT: Fast Vision Transformers with Hierarchical Attention (ICLR 2024 · similarity 0.89)
- Structured Initialization for Vision Transformers (NeurIPS 2025 · similarity 0.89)
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (AAAI 2024 · similarity 0.89)
- You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 2024 · similarity 0.88)
- ViT-CoMer: Vision Transformer with Convolutional Multi-scale Feature Interaction for Dense Predictions (CVPR 2024 · 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.
