The DOI matches its Crossref record; 34 of the 34 references it cites were found in a registry or an index; 4 of the 4 sources its claims cite were checked for support.
PTQ4ViT: Post-training Quantization for Vision Transformers with Twin Uniform Quantization
Open accessAbstract
The twin uniform quantization method is proposed to reduce the quantization error on activation values after softmax and GELU functions and to use a Hessian guided metric to evaluate different scaling factors.
Overview
CoChat Summary
TL;DR
PTQ4ViT is a post-training quantization framework for vision transformers that uses twin uniform quantization for post-softmax and post-GELU activations and a Hessian-guided metric to select scaling factors. It achieves near-lossless ImageNet classification accuracy at 8-bit quantization, with less than 0.5% top-1 drop on ViT, DeiT, and Swin models.
Key findings
- At 8-bit quantization, PTQ4ViT achieves less than 0.5% accuracy drop across tested models, e.g. ViT-B/384 drops 0.17% (86.00 → 85.82) and Swin-B/384 drops 0.04% (86.44 → 86.39), while base PTQ drops more than 1% on some vision transformers.
- At 6-bit quantization, PTQ4ViT reduces average accuracy drop to 2.1% versus 9.8% for base PTQ; e.g. DeiT-B/224 reaches 80.25% vs 81.80% FP32, and ViT-B/384 reaches 83.34% vs 86.00% FP32.
- PTQ4ViT outperforms EasyQuant and Liu et al. at 8-bit and 6-bit using only 32 calibration images versus 1024: DeiT-S/224 W8A8 79.47% vs 76.59% (EasyQuant) and 77.47% (Liu), and W6A6 76.28% vs 73.26% and 74.58%.
- Ablation shows both components matter: on ViT-S/224, using Hessian-guided metric alone improves 6-bit accuracy by 6.96%, and using twin uniform quantization without Hessian guidance drops 8-bit accuracy to 79.25%, even below base PTQ’s 80.47%.
Methods
- The paper proposes twin uniform quantization with two ranges R1 and R2 and separate scaling factors, using a data format where the most significant bit flags the range and constraining ∆R2 = 2^m ∆R1 so ranges can be aligned with shift operations on CPUs/GPUs.
- It proposes a Hessian-guided metric based on a diagonal Fisher information approximation to select scaling factors layer-by-layer, replacing MSE, cosine distance, and Pearson correlation coefficient, which the authors find inaccurate on vision transformers.
- Experiments use ImageNet classification with ViT, DeiT, and Swin models, 32 randomly selected calibration images, quantization of weights and inputs for fully connected layers and matrix multiplication inputs, and no quantization of softmax or normalization layers.
Limitations
- The paper states that at 4-bit quantization PTQ4ViT performs poorly: DeiT-B W4A4 reaches 60.91%, and with bias correction 64.39%, still lower than Liu et al.’s mixed-precision W4A4 result of 75.94%; it concludes mixed-precision is important for lower bit-widths.
- The paper states the Hessian-guided metric is closer to task loss than other metrics but still has a gap with task loss.
- Evident from the text: evaluation is limited to ImageNet classification on ViT/DeiT/Swin; the method assumes the floating-point network’s prediction approximates ground truth and uses a diagonal Fisher approximation, and no actual hardware latency or energy measurements are reported despite hardware-efficiency claims.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
4 Valid · 1 Not valid · 0 Not checked · 58 Own results and interpretations
Every source the paper cites for this claim was checked and supports it.
In the paper
“However, previous post-training quantization methods performed not well on vision transformer, resulting in more than 1% accuracy drop even in 8-bit quantization.”
- Also stated in: § 1 Introduction · p. 2 · plain
“Although PTQ has achieved great success on CNNs, directly bringing it to vision transformer results in more than 1% accuracy drop even with 8-bit quantization [16].”
- Also stated in: § 4.2 Results on ImageNet Classification Task · p. 11 · plain
“From this table, we observe that base PTQ results in more than 1% accuracy drop on some vision transformers even at the 8-bit quantization.”
What the sources say
- Supportedcites [16] Liu, Z., Wang, Y., Han, K., Zhang, W., Ma, S., Gao, W.: Post-training quantization for vision transformer, pp. 28092–28103 (2021). https://proceedings.neurips.cc/paper/2021/hash/ec8956637a99787bd197eacd77acce5e-Abstract.html source
Section 4.2 reports that conventional post-training quantization methods caused 8-bit top-1 accuracy losses of 5.82%, 3.21%, and 2.74% on DeiT-S. These results support the claim that the tested previous methods produced more than a 1% accuracy drop even at 8 bits. (LLM, conf=0.93)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § 1 Introduction · p. 2 · plain
Every source the paper cites for this claim was checked and supports it.
In the paper
“Based on the layer-wise reconstruction method in [14], the optimization can be approximated 6 by:”
What the sources say
- Supportedcites [14] Li, Y., et al.: BRECQ: pushing the limit of post-training quantization by block reconstruction. In: 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, 3–7 May 2021. OpenReview.net (2021). https://openreview.net/forum?id=POWv6hDd9XH source
The Introduction (p. 1) says that reconstructing each layer’s output can approximate quantization-induced task-loss degradation. The Preliminaries (p. 2) describe the resulting layer-by-layer feature-map reconstruction objective, while noting the assumptions under which it serves as a proxy. (LLM, conf=0.91)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Although bias correction [17] can improve the performance of PTQ4ViT, the result at 4-bit quantization is lower than the mixed-precision of Liu et al.”
What the sources say
- Partly supportedcites [17] Nagel, M., van Baalen, M., Blankevoort, T., Welling, M.: Data-free quantization through weight equalization and bias correction. In: 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), 27 October - 2 November 2019, pp. 1325–1334. IEEE (2019). https://doi.org/10.1109/ICCV.2019.00141 sourceExperimental — accuracy not yet measured
Section 5.1.2 reports that bias correction improves quantized MobileNetV2 performance, supporting the general benefit attributed to the method. The paper does not establish that it improves PTQ4ViT specifically at 4-bit quantization; its reported experiments concern different models and quantization settings. (LLM, conf=0.88)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“For instance, on ViT-S/224, the accuracy improvement is 0.46% at 8-bit while it is 6.96% at 6-bit.”
plainstates a causeSingle case- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Experiments show the quantized vision transformers achieve near-lossless prediction accuracy (less than 0.5% drop at 8-bit quantization) on the ImageNet classification task.”
- Also stated in: § 4.2 Results on ImageNet Classification Task · p. 11 · plain
“PTQ4ViT achieves less than 0.5% accuracy drop with 8-bit quantization.”
boosted- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § 4.2 Results on ImageNet Classification Task · p. 11 · plain
40 more claims
Create a free account to see all 63 claims and every source verdict.
- Probably not a substantial claim. 6
- It may state two separate claims as one. 14
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The cited source says what the claim attributes to it.
In the citing paper
“As shown in Fig.6, PTQ4ViT [42] introduces the concept of twin uniform quantization for post-Softmax activations. This method divides the activation values following Softmax into two distinct quantization ranges, denoted as R1 and R2. Each range is governed by a unique scaling factor, ∆ R 1 for R1 and ∆ R 2 for R2, respectively.”
From Model Quantization and Hardware Acceleration for Vision Transformers: A Comprehensive Survey (2024) · cites this paper as [42]
The cited source says what the claim attributes to it.
In the citing paper
“PTQ4ViT [42] applies twin uniform quantization to post-GELU activations, akin to its implementation in post-Softmax scenarios”
From Model Quantization and Hardware Acceleration for Vision Transformers: A Comprehensive Survey (2024) · cites this paper as [42]
The cited source says what the claim attributes to it.
In the citing paper
“PTQ4ViT [42] highlights the efficacy of a Hessian-guided metric for determining scaling factors, which can notably enhance the performance of quantized vision transformers, in contrast to other metrics that may not align with task-specific losses.”
From Model Quantization and Hardware Acceleration for Vision Transformers: A Comprehensive Survey (2024) · cites this paper as [42]
The cited source says what the claim attributes to it.
In the citing paper
“The experiments on ViT, DeiT, and Swin show that PTQ4ViT obtains near-lossless accuracy on the ImageNet dataset.”
From A Survey of Techniques for Optimizing Transformer Inference (2023) · cites this paper as [125]
The cited source says what the claim attributes to it.
In the citing paper
“Yuan et al. [125] observe that the activations after softmax and GELU operations differ from the traditional Gaussian distribution.”
From A Survey of Techniques for Optimizing Transformer Inference (2023) · cites this paper as [125]
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
§ 4.3 Ablation Study · p. 13
“| ViT-B/384 86.00 | ✓ | ✓ | ✓ | 85.89 | 83.19 |”
§ 4.1 Experiment Settings · p. 11
“| ViT-B/384 | 86.00 | 85.35(0.64) | 46.88(39.1) | 85.82(0.17) | 83.34(2.65) |”
Citations
Cites
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017) · arXiv · doi:10.48550/arxiv.1706.03762 · Cited by 120,404
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) · doi:10.18653/v1/n19-1423 · Cited by 54,325 · cited 1×
“Inspired by the success of transformers [6,2] on natu-ral language processing (NLP) tasks, researchers have brought the self-attention module into computer vision [7,15].”
MethodAn Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al. (2020) · arXiv · doi:10.48550/arxiv.2010.11929 · Cited by 47,472 · cited 3×
“ViT [7] is the first architecture that replaces all convolution layers, which achieves better results on image classification tasks.”
BackgroundMethodImageNet Large Scale Visual Recognition Challenge
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al. (2015) · International Journal of Computer Vision · doi:10.1007/s11263-015-0816-y · Cited by 39,380 · cited 1×
“We experiment on the ImageNet classification task [19].”
MethodSwin Transformer: Hierarchical Vision Transformer using Shifted Windows
Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., & Guo, B. (2021) · arXiv · Cited by 35,608
Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., & Guo, B. (2021) · 2021 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv48922.2021.00986 · Cited by 35,534 · cited 4×
“Vision transformers have been successfully applied to downstream tasks [15,24,12].”
BackgroundMethodInfluentialLanguage Models are Few-Shot Learners
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.14165 · Cited by 23,372
End-to-End Object Detection with Transformers
Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., & Zagoruyko, S. (2020) · Lecture Notes in Computer Science · doi:10.1007/978-3-030-58452-8_13 · Cited by 18,285 · cited 1×
“DETR [3] is the first work to replace the object detection head with a transformer, which directly regresses the bounding boxes and achieves comparable results with the CNN-based head.”
ResultEnd-to-End Object Detection with Transformers
Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., & Zagoruyko, S. (2020) · arXiv · doi:10.48550/arxiv.2005.12872 · Cited by 18,253
Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions
Wang, W., Xie, E., Li, X., Fan, D.-P., Song, K., Liang, D., Lu, T., Luo, P., & Shao, L. (2021) · 2021 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv48922.2021.00061 · Cited by 5,259 · cited 1×
“Following ViT, various vision transformer architectures have been proposed to boost performance [25,15].”
Background
Research works printed but not linked: 1
- Banner, R., Nahshan, Y., Soudry, D.: Post training 4-bit quantization of convolutional networks for rapid-deployment. In: Wallach, H.M., Larochelle, H., Beygelzimer, A., d’Alché-Buc, F., Fox, E.B., Garnett, R. (eds.) Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, 8–14 December 2019. Vancouver, BC, Canada, pp. 7948–7956 (2019). https://proceedings.neurips.cc/paper/2019/hash/c0a62e133894cdce435bcb4a5df1db2d-Abstract.html
Cited by
A Survey of Techniques for Optimizing Transformer Inference
Chitty-Venkata, K. T., Mittal, S., Emani, M., Vishwanath, V., & Somani, A. K. (2023) · Journal of Systems Architecture · doi:10.1016/j.sysarc.2023.102990 · Cited by 154
A Survey on Efficient Vision Transformers: Algorithms, Techniques, and Performance Benchmarking
Papa, L., Russo, P., Amerini, I., & Zhou, L. (2024) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2024.3392941 · Cited by 256
“Similarly to Liu et al. Yuan et al. [69] (2022) propose an effi-cientframeworkforpost-trainingquantizationnamedPTQ4ViT.”
Post-Training Quantization on Diffusion Models
Shang, Y., Yuan, Z., Xie, B., Wu, B., & Yan, Y. (2023) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr52729.2023.00196 · Cited by 414
“Other metrics, such as L1 distance, cosine distance, and KL divergence, can also be used to evaluate the distance of the tensors before and after quantization [19, 40].”
Self-citationRepQ-ViT: Scale Reparameterization for Post-Training Quantization of Vision Transformers
Li, Z., Xiao, J., Yang, L., & Gu, Q. (2023) · 2023 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv51070.2023.01580 · Cited by 127
“Although various PTQ methods for convolutional neural networks (CNNs) have been proposed in previous works with good performance, they produce disappointing results on ViTs, with more than 1% accuracy drop even in 8-bit quantization [35].”
MethodResultInfluentialPatch Similarity Aware Data-Free Quantization for Vision Transformers
Li, Z., Ma, L., Chen, M., Xiao, J., & Gu, Q. (2022) · Lecture Notes in Computer Science · doi:10.1007/978-3-031-20083-0_10 · Cited by 61
“In addition, several post-training quantization methods have been proposed to reduce the fine-tuning cost [9,17,22,24,27,32], including schemes for vision transformers [26,36], but they still require a small amount of real data for calibration and cannot achieve complete data-free.”
Background
Metadata
- Authors
- Yuan, Z., Xue, C., Chen, Y., Wu, Q., & Sun, G.
- Affiliations
- Peking University (CN)
- Corresponding author
- Sun, Guangyu
- Venue
- Lecture notes in computer science
- Publisher
- Springer Nature Switzerland
- Published
- 2022
- Pages
- 191-207
- ISSN
- 0302-9743 (ISSN-L)
- 1611-3349
- OpenAlex
- W4313069943
- Cited by
- 163
- References
- 29
- Citations by year
- 2026: 45 · 2025: 61 · 2024: 36 · 2023: 19 · 2022: 2
- FWCI
- 10.70
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Licence
- www.springernature.com/gp/researchers/text-and-data-mining (tdm)
- www.springernature.com/gp/researchers/text-and-data-mining (vor)
- Locations
- Lecture notes in computer science (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Funders
- National Natural Science Foundation of China (B18001, 62032001, 92064006, 61832020)
- Higher Education Discipline Innovation Project (B 18001)
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Data Compression Techniques
- Keywords
- post-training quantization, vision transformer, neural network compression, activation distribution, calibration, 8-bit quantization, ImageNet classification
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-28, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-28 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: PTQ4ViT: Post-training Quantization for Vision Transformers with Twin Uniform Quantization. https://app.cochat.ai/papers/10.1007/978-3-031-19775-8_12
- The paper
- Yuan, Z., Xue, C., Chen, Y., Wu, Q., & Sun, G. (2022). PTQ4ViT: Post-training Quantization for Vision Transformers with Twin Uniform Quantization. Lecture notes in computer science. https://doi.org/10.1007/978-3-031-19775-8_12
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.