Research map / Detection and segmentation / CNN features
Patch n’ Pack: NaViT, a Vision Transformer for any Aspect Ratio and Resolution
Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Peter Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim Alabdulmohsin, Avital Oliver, Piotr Padlewski and 3 more
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 NaViT packing with variable resolution and token dropping instead of fixed-resolution ViT, compute matched | ViT-B/32 to L/16 · JFT pretraining / ImageNet-A | accuracy: 14.7 -> 32.7 (B/32, 1.4x10^11) | Improved |
| NaViT vs ViT on ObjectNet with center crop at high compute | NaViT-B/16 · ObjectNet | accuracy: 52.8 -> 50.7 (B/16, 4.8x10^12) | Mixed |
| NaViT backbone instead of ViT for OWL-ViT detector | OWL-ViT-L/14 · LVIS | AP rare: 17.2 -> 24.3 | Improved |
| NaViT instead of ViT for contrastive pretraining | L/14 · ImageNet zero-shot | accuracy: 68.3 -> 72.9 | Improved |
| train with variable resolution U(64,Rmax) instead of fixed resolution (ablation) | NaViT-B/16 · JFT pretraining, ImageNet eval | accuracy | Improved |
| sample side length rather than area, with truncated normal biased to low resolution (ablation) | NaViT resolution sampling · JFT pretraining | accuracy | Improved |
| sample token drop rate from Beta distribution instead of constant (ablation) | NaViT-B/16 · JFT pretraining | accuracy | Improved |
| resolution-dependent token dropping rate (ablation) | NaViT · JFT pretraining | accuracy | Improved |
| decrease token dropping rate over training (ablation) | NaViT-B/16 · ImageNet 10-shot | accuracy | Improved |
| increase token dropping rate over training (ablation) | NaViT-B/16 · ImageNet 10-shot | accuracy | Hurt |
| use factorized additive positional embeddings instead of learned 2D (Pix2struct) or ViT 1D (ablation) | NaViT-B/16 · JFT, evaluated at new resolutions | accuracy | Improved |
| fine-tune NaViT at variable resolution 64:512 instead of single resolution (ablation) | NaViT-B/16 · ImageNet-1k | accuracy | No clear effect |
Closest papers by meaning
- ResFormer: Scaling ViTs With Multi-Resolution Training (CVPR 2023 · similarity 0.89)
- VITRIX-UniViTAR: Unified Vision Transformer with Native Resolution (NeurIPS 2025 · similarity 0.89)
- Native-Resolution Image Synthesis (NeurIPS 2025 · similarity 0.89)
- FlexiViT: One Model for All Patch Sizes (CVPR 2023 · similarity 0.89)
- HIRI-ViT: Scaling Vision Transformer With High Resolution Inputs (TPAMI 2024 · similarity 0.88)
- NViST: In the Wild New View Synthesis from a Single Image with Transformers (CVPR 2024 · 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.
