Research map / Architectures and efficiency / Quantization
Q-ViT: Accurate and Fully Quantized Low-bit Vision Transformer
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 |
|---|---|---|---|
| add Information Rectification Module (IRM) (ablation) | fully quantized DeiT-S baseline · ImageNet 2-bit | Top-1 accuracy (%): 68.2 -> 69.9 | Helped |
| add Distribution Guided Distillation (DGD) (ablation) | fully quantized DeiT-S baseline · ImageNet 2-bit | Top-1 accuracy (%): 68.2 -> 70.5 | Helped |
| combine IRM and DGD (ablation) | fully quantized DeiT-S baseline · ImageNet 2-bit | Top-1 accuracy (%): 68.2 -> 72.0 | Helped |
| add IRM (ablation) | fully quantized DeiT-S baseline · ImageNet 4-bit | Top-1 accuracy (%): 79.7 -> 80.2 | Helped |
| add DGD (ablation) | fully quantized DeiT-S baseline · ImageNet 4-bit | Top-1 accuracy (%): 79.7 -> 80.4 | Helped |
| add IRM and DGD (ablation) | fully quantized DeiT-S baseline · ImageNet 3-bit | Top-1 accuracy (%): 77.8 -> 79.0 | Helped |
| Q-ViT vs baseline | DeiT-B 2-bit · ImageNet | Top-1 accuracy (%): 70.4 -> 74.2 | Helped |
| Q-ViT vs baseline | Swin-S 2-bit · ImageNet | Top-1 accuracy (%): 72.7 -> 76.9 | Helped |
| Q-ViT vs LSQ | DeiT-S 4-bit · ImageNet | Top-1 accuracy (%): 79.6 -> 80.9 | Helped |
| baseline 4-bit vs LSQ | DeiT-S · ImageNet | Top-1 accuracy (%): 79.6 -> 79.7 | No clear effect |
| quantize query, key, value and attention weight to 2 bits (vs MLP or linear weights) (ablation) | DeiT 2-bit analysis · ImageNet | Top-1 accuracy drop (%): 10.03% drop vs 1.78% (MLP) and 4.26% (linear weights) | Hurt |
Closest papers by meaning
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (AAAI 2024 · similarity 0.92)
- RPFQ-ViT: Rotated Phase-Frame Quantization for Extremely Low-Bit Weights in Vision Transformers (NeurIPS 2026 · similarity 0.92)
- GradQ-ViT: Robust and Efficient Gradient Quantization for Vision Transformers (AAAI 2025 · similarity 0.92)
- GPLQ: A General, Practical, and Lightning QAT Method for Vision Transformers (NeurIPS 2025 · similarity 0.90)
- AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers (AAAI 2025 · similarity 0.90)
- MimiQ: Low-Bit Data-Free Quantization of Vision Transformers with Encouraging Inter-Head Attention Similarity (AAAI 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.