Research map / Architectures and efficiency / Quantization
UQ-ViT: Harmonizing Extreme Activations with Hardware-Friendly Uniform Quantization in Vision Transformers
Tao Jiang, Yucheng Jiang, Xiwen Yao, Gong Cheng, Junwei Han
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 DeMax and NormQuant to uniform-quantized RepQ-ViT baseline (ablation) | RepQ-ViT with uniform quantization · ImageNet W4/A4, DeiT-S | top-1 accuracy: 68.98 -> 72.13 | Improved |
| add DeMax and NormQuant to uniform-quantized RepQ-ViT baseline (ablation) | RepQ-ViT with uniform quantization · ImageNet W4/A4, ViT-B | top-1 accuracy: 67.61 -> 72.07 | Improved |
| add NormQuant alone (ablation) | uniform-quantized baseline · ImageNet W4/A4, ViT-B | top-1 accuracy: 67.61 -> 70.83 | Improved |
| add DeMax alone (ablation) | uniform-quantized baseline · ImageNet W4/A4, ViT-B | top-1 accuracy: 67.61 -> 68.30 | Improved |
| add DeMax on top of NormQuant (ablation) | UQ-ViT · ImageNet W4/A4, DeiT-T | top-1 accuracy: 58.65 -> 59.71 | Improved |
| apply NormQuant to fc1 (ablation) | UQ-ViT · ImageNet W4/A4, DeiT-S | top-1 accuracy: 57.63 -> 72.13 (+6.99% stated for fc1) | Improved |
| apply NormQuant to qkv (ablation) | UQ-ViT · ImageNet W4/A4, DeiT-S | top-1 accuracy: 61.02 -> 72.13 (+5.61% stated for qkv) | Improved |
| apply NormQuant to fc2 (ablation) | UQ-ViT · ImageNet W4/A4, DeiT-S | top-1 accuracy: 69.92 -> 72.13 (+1.92% stated for fc2) | Improved |
| apply NormQuant to proj (ablation) | UQ-ViT · ImageNet W4/A4, DeiT-S | top-1 accuracy: 72.01 -> 72.13 (+0.30% stated for proj) | No clear effect |
| use uniform-only UQ-ViT instead of non-uniform ERQ at 6-bit | UQ-ViT · ImageNet W6/A6 | top-1 accuracy: DeiT-S 79.07 vs ERQ 79.03; Swin-B 84.99 vs ERQ 85.02 | No clear effect |
| UQ-ViT (uniform) versus ERQ at 4-bit | UQ-ViT · ImageNet W4/A4 | top-1 accuracy: ViT-B 72.07 vs ERQ 76.63; Swin-S 79.93 vs ERQ 80.74 | Mixed |
| apply UQ-ViT to Mask R-CNN Swin-T at 4-bit | Mask R-CNN Swin-T · COCO W4/A4 | APbox: 36.8 -> 39.0 (vs ERQ) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- RepQ-ViT — RepQ-ViT: Scale Reparameterization for Post-Training Quantization of Vision Transformers (ICCV 2023) · better on 3 of 3 shared results
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)
- GPLQ: A General, Practical, and Lightning QAT Method for Vision Transformers (NeurIPS 2025 · similarity 0.91)
- GradQ-ViT: Robust and Efficient Gradient Quantization for Vision Transformers (AAAI 2025 · similarity 0.90)
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (AAAI 2024 · similarity 0.90)
- APHQ-ViT: Post-Training Quantization with Average Perturbation Hessian Based Reconstruction for Vision Transformers (CVPR 2025 · similarity 0.90)
- RPFQ-ViT: Rotated Phase-Frame Quantization for Extremely Low-Bit Weights in Vision Transformers (NeurIPS 2026 · similarity 0.89)
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.
