Research map / Architectures and efficiency / Attention and transformers
FlexiViT: One Model for All Patch Sizes
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 |
|---|---|---|---|
| randomize patch size during training and resize embedding parameters for each size | ViT · ImageNet-21k and downstream vision tasks | performance | Helped |
| use PI-resize instead of bilinear resizing for patch-embedding weights (ablation) | FlexiViT · ImageNet-21k classification | accuracy | Helped |
| initialize the student from a teacher ViT and distill its predictions (ablation) | FlexiViT-B · ImageNet-21k classification | accuracy | Helped |
| use triangular rather than uniform patch-size sampling (ablation) | FlexiViT · ImageNet-21k classification | performance | No clear effect |
| vary the underlying patch-embedding parameter shape (ablation) | FlexiViT · ImageNet-21k classification | performance | No clear effect |
| vary the underlying position-embedding parameter shape (ablation) | ViT-S/16 · ImageNet-1k classification | training performance | No clear effect |
| learn separate patch-embedding weights for each patch size (ablation) | FlexiViT · ImageNet-21k classification | performance | Hurt |
| use a patch-size curriculum, starting with large patches and switching to the target size (ablation) | ViT-B/16 pretraining · ImageNet-21k | compute at matched accuracy | Helped |
| increase the fraction of training assigned to the larger patch size (ablation) | FasterViT · ImageNet-21k pretraining | compute | Helped |
| vary patch embedding stride instead of patch size (ablation) | ViT · ImageNet-21k classification | compute-accuracy tradeoff | Mixed |
| vary model depth by attaching a shared head to intermediate layers (ablation) | ViT · ImageNet-21k classification | compute-accuracy tradeoff | Hurt |
| ensemble predictions across patch-size scales instead of using one scale at comparable compute (ablation) | FlexiViT-B · ImageNet-21k | accuracy per compute | Hurt |
Closest papers by meaning
- Scaling Laws in Patchification: An Image Is Worth 50,176 Tokens And More (ICML 2025 · similarity 0.89)
- Slicing Vision Transformer for Flexible Inference (NeurIPS 2024 · similarity 0.89)
- Elastic ViTs from Pretrained Models without Retraining (NeurIPS 2025 · similarity 0.88)
- Linearly Decomposing and Recomposing Vision Transformers for Diverse-Scale Models (NeurIPS 2024 · similarity 0.88)
- Your ViT is Secretly an Image Segmentation Model (CVPR 2025 · similarity 0.88)
- REOrdering Patches Improves Vision Models (NeurIPS 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.