Research map / Architectures and efficiency / Pruning and efficient tuning
Getting ViT in Shape: Scaling Laws for Compute-Optimal Model Design
Ibrahim Alabdulmohsin, Xiaohua Zhai, Alexander Kolesnikov, Lucas Beyer
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 compute-optimal shape SoViT-400m/14 instead of ViT-L/16 at 384px | fine-tuning · ImageNet v2 | accuracy: 80.4 -> 83.2 | Improved |
| use SoViT-400m/14 instead of ViT-g/14 at 518px (2.4x smaller) | fine-tuning · ImageNet Val | accuracy: 90.2 -> 90.3 | No clear effect |
| use SoViT-400m/14 instead of ViT-L/16 in linear probe | linear probe · ImageNet-A | accuracy: 67.8 -> 76.4 | Improved |
| use SoViT-400m/14 instead of ViT-g/14 in linear probe | linear probe · ImageNet-R | accuracy: 90.3 -> 89.0 | Hurt |
| use SoViT-400m/14 instead of ViT-g/14 | 10-shot linear transfer · ImageNet | accuracy: 84.0 -> 84.1 | No clear effect |
| use SoViT-400m/14 instead of ViT-L/16 | 10-shot linear transfer · CIFAR100 | accuracy: 82.2 -> 86.7 | Improved |
| use SoViT-400m/14 instead of ViT-g/14 | 10-shot linear transfer · Birds | accuracy: 88.5 -> 88.8 | No clear effect |
| use SoViT-400m/14 instead of ViT-L/16 in LiT | zero-shot ImageNet · accuracy | accuracy: 79.9 -> 82.2 | Improved |
| use SoViT-400M instead of ViT-g/14 in multitask decoding | OCR-VQA · accuracy [%] | accuracy: 52.5 -> 52.9 | Improved |
| use SoViT-400M instead of ViT-g/14 in multitask decoding | GQA · accuracy [%] | accuracy: 58.0 -> 56.0 | Hurt |
| use SoViT-400M instead of ViT-g/14 in multitask decoding | COCO Captioning · CIDEr | CIDEr: 126 -> 125 | No clear effect |
| use SoViT-400m/14 instead of ViT-g/14 as UViM encoder | COCO panoptic · PQ | PQ: 44.8 -> 43.7 | Hurt |
Closest papers by meaning
- Scaling Vision Transformers (CVPR 2022 · similarity 0.91)
- Auto-scaling Vision Transformers without Training (ICLR 2022 · similarity 0.90)
- DeepCompress-ViT: Rethinking Model Compression to Enhance Efficiency of Vision Transformers at the Edge (CVPR 2025 · similarity 0.87)
- Slicing Vision Transformer for Flexible Inference (NeurIPS 2024 · similarity 0.87)
- Scaling Laws in Patchification: An Image Is Worth 50,176 Tokens And More (ICML 2025 · similarity 0.87)
- Vision Transformers Beat WideResNets on Small Scale Datasets Adversarial Robustness (AAAI 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.
