Research map / Architectures and efficiency / Quantization
BLOB-Q: Boosting Low Bit ViT Quantization via Global Optimization on Model Distortion
Wang Mark, Kaixin Xu, Xue Geng, Fen Fang, Mohamed Aly, Xulei Yang, Min Wu, Weisi Lin
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 |
|---|---|---|---|
| optimize model-level distortion instead of layer distortion (greedy local solver) (ablation) | BLOB-Q · ImageNet val ViT-S 4-bit | top-1 accuracy: 0.11% -> 20.52% | Improved |
| replace the local solver with the global-optimal DP solver (ablation) | BLOB-Q · ImageNet val ViT-S 4-bit | top-1 accuracy: 20.52% -> 76.85% | Improved |
| replace the local solver with the global-optimal DP solver (ablation) | BLOB-Q · ImageNet val ViT-B 4-bit | top-1 accuracy: 1.25% -> 82.84% | Improved |
| replace the local solver with the global-optimal DP solver (ablation) | BLOB-Q · ImageNet val DeiT-B/384 4-bit | top-1 accuracy: 49.92% -> 81.64% | Improved |
| replace the local solver with the global-optimal DP solver (ablation) | BLOB-Q · ImageNet val ViT-S 6-bit | top-1 accuracy: 62.87% -> 80.76% | Improved |
| optimize model distortion instead of layer distortion (local solver) (ablation) | BLOB-Q · ImageNet val DeiT-B 8-bit | top-1 accuracy: 45.81% -> 80.81% | Improved |
| increase calibration set from 32 to 512 images (ablation) | BLOB-Q · ImageNet val DeiT-S 4-bit | top-1 accuracy: 75.33% -> 76.13% | Improved |
| increase calibration set from 512 to 1024 images (ablation) | BLOB-Q · ImageNet val DeiT-S 6-bit | top-1 accuracy: 79.65% -> 79.42% | No clear effect |
| add BLOB-Q mixed-precision allocation on top of ERQ | ERQ · ImageNet val ViT-B 4-bit | top-1 accuracy: 76.63 -> 80.14 | Improved |
| add BLOB-Q mixed-precision allocation on top of ERQ | ERQ · ImageNet val DeiT-S 4-bit | top-1 accuracy: 72.56 -> 77.87 | Improved |
| add BLOB-Q mixed-precision allocation on top of ERQ | ERQ · ImageNet val DeiT-B 6-bit | top-1 accuracy: 81.41 -> 82.03 | Improved |
| restrict per-layer bit-width to {4,8} instead of [1,10] (ablation) | BLOB-Q · ImageNet val ViT-B 6 MP | top-1 accuracy: 85.07 -> 84.26 | Hurt |
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.
- PTQ4ViT — PTQ4ViT: Post-Training Quantization for Vision Transformers with Twin Uniform Quantization (ECCV 2022) · 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
- GPLQ: A General, Practical, and Lightning QAT Method for Vision Transformers (NeurIPS 2025 · similarity 0.91)
- 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 0.90)
- GradQ-ViT: Robust and Efficient Gradient Quantization for Vision Transformers (AAAI 2025 · similarity 0.90)
- AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers (AAAI 2025 · similarity 0.90)
- ERQ: Error Reduction for Post-Training Quantization of Vision Transformers (ICML 2024 · 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.
