Research map / Architectures and efficiency / Quantization
Post-training quantization of ViTs
Reduces quantization error in vision transformers using Hessian, Fisher, and outlier-aware reconstruction.
105 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2020–2026) · an approach in Quantization, keyword cluster “post quantization · ptq · compression”
State History Tried Baselines Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
26 papers in 2023–24 and 70 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 22.0% to 20.5%.
Open questions the data points to
- Papers report QDrop on ImageNet, DeiT-T, W3/A3 (top-1 accuracy) anywhere from 15.68 to 46.88: comparisons against it depend on whose number you take.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2020 · Bi3D: Stereo Depth Estimation via Binary Classifications (CVPR 2020 · 88 citations)
- 2021 · Post-Training Quantization for Vision Transformer (NeurIPS 2021 · 3 citations)
- 2022 · QDrop: Randomly Dropping Quantization for Extremely Low-bit Post-Training Quantization (ICLR 2022 · 43 citations)
- 2023 · PD-Quant: Post-Training Quantization Based on Prediction Difference Metric (CVPR 2023 · 75 citations)
- 2024 · Make RepVGG Greater Again: A Quantization-Aware Approach (AAAI 2024 · 33 citations)
- 2025 · GradQ-ViT: Robust and Efficient Gradient Quantization for Vision Transformers (AAAI 2025 · 16 citations)
- 2026 · LampQ: Towards Accurate Layer-wise Mixed Precision Quantization for Vision Transformers (AAAI 2026 · 3 citations)
Earliest papers here
- Bi3D: Stereo Depth Estimation via Binary Classifications (CVPR 2020 · 88 citations)
- Position-based Scaled Gradient for Model Quantization and Pruning (NeurIPS 2020 · 17 citations)
- Post-Training Quantization for Vision Transformer (NeurIPS 2021 · 3 citations)
What has been tried
56 papers here report 633 measured changes (a component added, removed or swapped, compared with a baseline): helped 72%, no clear effect 6%, hurt 20%, mixed 2%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Replace custom L2 with normal L2 | ImageNet validation · INT8 top-1 accuracy | 50.3 -> 61.6 | Helped | Make RepVGG Greater Again: A Quantization-Aware Approach (AAAI 2024 · 33 citations) |
| Increase group count from 8 to 12 | ImageNet validation, 4/4-bit · top-1 accuracy | 72.99 -> 73.61 | Helped | Instance-Aware Group Quantization for Vision Transformers (CVPR 2024 · 20 citations) |
| Add IQS to ALA | ImageNet, DeiT-Tiny · accuracy | 65.62 -> 69.48 | Helped | GradQ-ViT: Robust and Efficient Gradient Quantization for Vision Transformers (AAAI 2025 · 16 citations) |
| Remove VAE context input | COCO Panoptic · PQ | 43.4±0.0008 -> 39.0±0.0023 | Hurt | Finite Scalar Quantization: VQ-VAE Made Simple (ICLR 2024 · 13 citations) |
| Add image-wise and layer-wise bit mapping, bit-aware clipping, and calibration fine-tuning | Urban100, 4-bit EDSR · PSNR | 24.27 -> 25.11 | Helped | AdaBM: On-the-Fly Adaptive Bit Mapping for Image Super-Resolution (CVPR 2024 · 11 citations) |
| Add CSB to basic linear quantization | ImageNet 256×256, 250 steps, W4A8 · FID | 22.54 -> 8.17 | Helped | PTQ4DiT: Post-training Quantization for Diffusion Transformers (NeurIPS 2024 · 7 citations) |
| Increase candidate assignment set length from n=1 to n=2 | ImageNet 256×256, 2-bit, 50 steps · FID | 60.16 -> 12.42 | Helped | VQ4DiT: Efficient Post-Training Vector Quantization for Diffusion Transformers (AAAI 2025 · 6 citations) |
| Use APHQ-ViT instead of QDrop | ImageNet, W3/A3 · top-1 accuracy | 38.31 -> 59.11 | Helped | APHQ-ViT: Post-Training Quantization with Average Perturbation Hessian Based Reconstruction for Vision Transformers (CVPR 2025 · 6 citations) |
| Add AILoC | ImageNet, DeiT-T, W3/A3 · top-1 accuracy | 23.28 -> 38.51 | Helped | AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers (AAAI 2025 · 5 citations) |
| Add diagonal and low-rank FIM components together (DPLR-FIM) | ImageNet, W3/A3 across ViT architectures · top-1 accuracy | 64.09 vs 60.02; 77.63 vs 76.29; 55.55 vs 55.54; 69.13 vs 68.68; 76.54 vs 76.32; 77.26 vs 75.08; 78.82 vs 77.87 | Helped | FIMA-Q: Post-Training Quantization for Vision Transformers by Fisher Information Matrix Approximation (CVPR 2025 · 5 citations) |
| Add inter-head coherency loss to synthesis | ImageNet · accuracy | 13.28 -> 42.99 | Helped | MimiQ: Low-Bit Data-Free Quantization of Vision Transformers with Encouraging Inter-Head Attention Similarity (AAAI 2025 · 4 citations) |
| add scale-aware LoRA optimization to QALoRA | ImageNet 256×256, W4A4 · FID | 11.42 -> 9.55 | Helped | EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models (ICLR 2024 · 4 citations) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| QDrop | ImageNet, DeiT-T, W3/A3 | top-1 accuracy | 15.68 – 46.88 | 2 papers, e.g. AAAI 2025 and CVPR 2025 |
What reviewers will expect
Datasets and baselines named in the 37 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- ImageNet · 13 of 37 papers
- COCO · 5 of 37 papers
- ImageNet-1k · 4 of 37 papers
- CIFAR10 · 4 of 37 papers
- CIFAR100 · 4 of 37 papers
- MSCOCO · 2 of 37 papers
- WikiText2 · 2 of 37 papers
- HellaSwag · 2 of 37 papers
Baselines
- RepQ-ViT · 8 of 37 papers
- PTQ4ViT · 7 of 37 papers
- RTN · 7 of 37 papers
- APQ-ViT · 6 of 37 papers
- GPTQ · 6 of 37 papers
- SmoothQuant · 6 of 37 papers
- SVDQuant · 5 of 37 papers
- QDrop · 4 of 37 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 96 papers; from OpenAlex and the papers' affiliations).
- Google · 5 papers
- Huawei · 5 papers
- ByteDance · 4 papers
- Seoul National University · 3 papers
- Nanjing University · 3 papers
- Xiamen University · 3 papers
- Tsinghua University · 3 papers
- Shanghai Jiao Tong University · 3 papers
- Sogang University · 3 papers
- Seoul National University of Science and Technology · 2 papers
Use it
Download BibTeX of 32 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Post-training weight quantization (100 papers)
- Post-training quantization of LLMs (136 papers)
- KV cache quantization (72 papers)
Elsewhere in the topic
- Mixed-precision quantization (92 papers)
- Low-precision training formats (58 papers)
- Ultra-low-bit LLM quantization (56 papers)
- Binary neural networks (27 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).