Research map / Architectures and efficiency / Quantization
PTQ4ViT: Post-Training Quantization for Vision Transformers with Twin Uniform Quantization
Zhihang Yuan, Chenhao Xue, Yiqi Chen, Qiang Wu, Guangyu Sun
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 |
|---|---|---|---|
| add Hessian guided metric to base PTQ (ablation) | ViT-S/224 W6A6 · ImageNet-1k val | top-1 accuracy: 70.24 -> 77.20 | Improved |
| add Hessian guided metric to base PTQ (ablation) | ViT-S/224 W8A8 · ImageNet-1k val | top-1 accuracy: 80.47 -> 80.93 | Improved |
| add Hessian guided metric to base PTQ (ablation) | ViT-B/384 W6A6 · ImageNet-1k val | top-1 accuracy: 46.89 -> 79.99 | Improved |
| add Hessian guided metric to twin quantization (ablation) | ViT-S/224 W8A8 · ImageNet-1k val | top-1 accuracy: 79.25 -> 81.00 | Improved |
| add Hessian guided metric to twin quantization (ablation) | ViT-S/224 W6A6 · ImageNet-1k val | top-1 accuracy: 74.07 -> 78.63 | Improved |
| add twin uniform quantization (softmax and GELU) on top of Hessian metric (ablation) | ViT-S/224 W6A6 · ImageNet-1k val | top-1 accuracy: 77.20 -> 78.63 | Improved |
| add twin uniform quantization on top of Hessian metric (ablation) | ViT-B/384 W6A6 · ImageNet-1k val | top-1 accuracy: 79.99 -> 83.19 | Improved |
| add twin uniform quantization without Hessian metric (ablation) | ViT-S/224 W8A8 · ImageNet-1k val | top-1 accuracy: 80.47 -> 79.25 | Hurt |
| use PTQ4ViT instead of base PTQ | ViT-B/384 W6A6 · ImageNet-1k val | top-1 accuracy: 46.88 -> 83.34 | Improved |
| use PTQ4ViT instead of Liu et al. (MP) | DeiT-B W4A4 · ImageNet-1k val | top-1 accuracy: 75.94 -> 60.91 | Hurt |
| add bias correction to PTQ4ViT | DeiT-B W4A4 · ImageNet-1k val | top-1 accuracy: 60.91 -> 64.39 | Improved |
| use PTQ4ViT instead of EasyQuant with fewer calibration images (32 vs 1024) | DeiT-S W8A8 · ImageNet-1k val | top-1 accuracy: 76.59 -> 79.47 | Improved |
Later papers that compared with this method
8 later accepted papers report PTQ4ViT as a baseline in their results tables; 7 of them report a better number on most of the results they share with it.
- BLOB-Q: Boosting Low Bit ViT Quantization via Global Optimization on Model Distortion (ECCV 2026 · better on 1 of 1 shared result)
- Quantization without Tears (CVPR 2025 · better on 1 of 1 shared result)
- LRA-QViT: Integrating Low-Rank Approximation and Quantization for Robust and Efficient Vision Transformers (ICML 2025 · better on 0 of 1 shared result)
- PTMQ: Post-training Multi-Bit Quantization of Neural Networks (AAAI 2024 · better on 3 of 4 shared results)
- Bit-Shrinking: Limiting Instantaneous Sharpness for Improving Post-Training Quantization (CVPR 2023 · better on 4 of 4 shared results)
- NoisyQuant: Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers (CVPR 2023 · better on 4 of 4 shared results)
- RepQ-ViT: Scale Reparameterization for Post-Training Quantization of Vision Transformers (ICCV 2023 · better on 2 of 2 shared results)
- Jumping through Local Minima: Quantization in the Loss Landscape of Vision Transformers (ICCV 2023 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers (AAAI 2025 · similarity 0.93)
- APHQ-ViT: Post-Training Quantization with Average Perturbation Hessian Based Reconstruction for Vision Transformers (CVPR 2025 · similarity 0.91)
- Instance-Aware Group Quantization for Vision Transformers (CVPR 2024 · similarity 0.91)
- ERQ: Error Reduction for Post-Training Quantization of Vision Transformers (ICML 2024 · similarity 0.91)
- FIMA-Q: Post-Training Quantization for Vision Transformers by Fisher Information Matrix Approximation (CVPR 2025 · similarity 0.90)
- GPLQ: A General, Practical, and Lightning QAT Method for Vision Transformers (NeurIPS 2025 · similarity 0.90)
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.
