REVIEW 3 major objections 5 minor 45 references
Zero-shot Quantization: A Comprehensive Survey
T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash
Pith's one-line read Without any real data, zero-shot quantization compresses neural networks; this survey sorts the field into three families.
desk verdict A useful first ZSQ survey with a sound taxonomy, but its comparison table mixes incompatible settings and needs correction before the quantitative claims can be trusted. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the three-way taxonomy of data acquisition in zero-shot quantization: synthesis-free methods, which recover accuracy without any generated data; generator-based methods, which train a GAN-style generator to produce synthetic examples; and noise-optimization-based methods, which directly optimize input noise through a two-step pattern of dataset synthesis followed by model quantization. Each family splits further into zero-shot quantization-aware training (QAT) and zero-shot post-training quantization (PTQ). The taxonomy does the work of making heterogeneous methods comparable and exposing shared baseline losses, such as batch-normalization statistics loss for CNNs and patch similarity entropy loss for vision transformers.
What would settle it
Re-run three representative methods, one from each family, under a single fixed protocol: the same ResNet-18 model, the same W4A4 bit-width, the same number of synthetic images, and the same evaluation split. If the accuracy ordering differs from the ordering implied by Table 1, the survey's comparative conclusions depend on the heterogeneous settings rather than on the methods themselves.
Extended reading notes
Core claim
The organizing discovery is that the entire ZSQ literature can be classified by data-generation strategy rather than by quantization technique. Synthesis-free methods never create data; they repair quantization damage using structural properties such as cross-layer weight equalization, Hessian approximations, or closed-form adjustment after pruning. Generator-based methods train a separate generative model to produce synthetic images for fine-tuning or calibration. Noise-optimization-based methods skip the generator and directly optimize random noise, usually by matching batch-normalization statistics and class predictions of the pre-trained model, with post-training variants needing as few as 32 images for vision transformers. The paper presents this taxonomy as its principal contribution, alongside a unified problem statement and a performance comparison table.
Load-bearing premise
The comparison table's ability to rank methods assumes that accuracy numbers taken from different papers are directly comparable, even though the papers use different models, bit-widths, and synthetic-image budgets, so relative standing could change under a uniform protocol.
Editorial extensions
If this is right
- A newcomer can identify which ZSQ family a method belongs to and which baseline losses it builds on, without reading every primary paper.
- The comparison table shows noise-optimization-based PTQ and synthesis-free methods reaching accuracy close to generator-based QAT while using far fewer synthetic images.
- Because BNS loss anchors CNN methods while ViT methods need attention-based losses like patch similarity entropy, architecture-specific ZSQ development will continue.
- The paper's named future directions include theoretical bounds on data-free accuracy, faster synthetic-data generation, and combining quantization with pruning or low-rank compression.
Reading between the lines
- If the taxonomy is right, the practical bottleneck of ZSQ is synthetic-data quality rather than quantization itself, so advances in diffusion-based image generation should transfer directly to ZSQ performance.
- The comparison table mixes CNN and ViT results, different bit-widths, and different synthetic-data budgets, so cross-method rankings are suggestive only; a standardized benchmark with fixed model, bit-width, and sample budget would settle which family is best.
- The paper's image-classification focus suggests a ready test of the taxonomy's generality: zero-shot quantization of language models, which lack batch-normalization statistics, may force a fourth family or a new baseline loss.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This manuscript surveys zero-shot quantization (ZSQ), a setting in which a pre-trained network is quantized without access to real training data. It formalizes the ZSQ problem, identifies three main challenges, and organizes existing methods into a three-way taxonomy: synthesis-free, generator-based, and noise-optimization-based approaches. For each method, the paper summarizes the motivation, core idea, and reported findings, and it provides a comparison table intended to serve as a performance baseline. The descriptive and taxonomic content is largely coherent and appears to align with the cited papers, but the quantitative comparison in Table 1 mixes incompatible evaluation settings and contains labeling inconsistencies. As presented, the table cannot support the paper's comparative-performance claims, which weakens one of the stated contributions.
Significance. If the descriptive content is taken on its own terms, this survey is a useful entry point for researchers entering ZSQ: it provides a formal problem definition, a clear and sensible three-way categorization with a helpful schematic (Figure 1), detailed summaries of representative methods, and a reasonable set of future directions. The paper also attempts a quantitative baseline table, and that part currently fails because the rows are not commensurable: ResNet-18 W4A4 numbers are mixed with DeiT-Tiny W8A8/W4A8 numbers, and the '# Images' column conflates totals, per-class counts, and calibration-set sizes. The 'first in-depth survey' claim is plausible but would be stronger with an explicit statement of search scope and inclusion criteria. Self-citations such as SynQ appear in the table, but the taxonomic and descriptive claims do not depend on those self-cited results, so circularity is not a serious concern.
major comments (3)
- [Table 1 and Section 3] The cross-method comparison in Table 1 is not valid as presented. The caption states that the table reports W4A4/W3A3 accuracy for ResNet-18 on ImageNet, yet the rows for PSAQ-ViT, SMI, and CLAMP-ViT report W8A8/W4A8 accuracy for DeiT-Tiny via footnotes. The '# Images' column also mixes total generator-training sample counts (1.28M), total synthesized-sample counts (5.12K), calibration-set sizes (1K), and ViT-specific set sizes (32). Because these quantities are not commensurable, any ranking or 'compare their performance' conclusion drawn from Table 1, as advertised in the Introduction and in Section 3, is not supported. I recommend restructuring the table into per-setting blocks, clearly stating the source of every number (original paper versus re-implementation), or removing the quantitative comparison and presenting the table as a qualitative feature summary only.
- [Table 1, DeepInversion row] The DeepInversion row carries the footnotes * and dagger, which are defined only for W8A8/W4A8 accuracy of DeiT-Tiny. DeepInversion is described in Section 6.1 as a CNN-oriented method with ResNet-18 experiments, so attaching DeiT-Tiny footnotes to it is misleading and makes the reported 70.27/64.28 numbers ambiguous. The row should either be relabeled with the actual evaluation protocol or removed from the footnoted group. The same ambiguity affects the GenQ and MixMix rows, where the annotation '1K (section sign)' is explained only as 'PTQ setting'; please clarify whether 1K is the calibration-set size or the size of the synthetic dataset, especially for GenQ, which Section 5.2 describes as using diffusion-based synthesis.
- [Section 3, evaluation methodology] The manuscript states that it evaluates ResNet-18 accuracy 'for fair benchmarking across methods' but does not say whether the numbers were taken from the original papers or produced by a unified re-implementation, nor which quantization configuration was used (per-channel versus per-tensor, symmetric versus asymmetric, calibration protocol, etc.). Without this information, even internally consistent rows cannot be independently checked, and the fairness claim in Section 3 is not verifiable. Please add a methodology paragraph or detailed table footnote describing the exact protocol, including how the ViT rows were obtained and why a ResNet-18 header applies to them.
minor comments (5)
- [Section 2.1, heading] The heading 'QA T and PTQ' contains an unintended space and should read 'QAT and PTQ'.
- [Section 6.2, first sentence] The statement that PTQ methods are 'typically evaluated with sets of 1,000 for CNNs and 32 for ViTs' is difficult to reconcile with Table 1, where noise-optimization-based QAT rows show 5.12K images; please clarify which numbers are calibration-set sizes and which are totals.
- [References and in-text citations] In-text citation formats are inconsistent, for example 'GenQ[2024b]' and 'AdaSG[2023b]' versus '[Li et al., 2024b]'; please standardize to a single style.
- [Section 7, 'Faster generation of synthetic datasets'] The estimate '1 to 4 RTX 4090 GPU hours' for generating 5,120 images should specify the method and exact setting to which it refers.
- [Table 1 caption] The notation 'WBAB' in the caption is used without definition; please spell out that W4A4 means weights and activations quantized to 4 bits, and clarify whether the notation applies only to weight-activation quantization.
Circularity Check
No circular derivation: the survey's taxonomy and performance table are descriptive compilations; self-citations are literature references, not load-bearing inputs.
full rationale
This paper is a survey, not a derivation. Its central claims are that it provides the first in-depth ZSQ overview and a taxonomy of methods (synthesis-free, generator-based, and noise-optimization-based). Those claims are supported by citing and describing the surveyed literature, not by deriving any quantity from the survey's own output. The only self-references are in the bibliography and in Table 1: SynQ [Kim et al., 2025b] is included as one of the surveyed methods, and earlier U Kang group papers are cited in the introduction and future-work sections (e.g., [Kim et al., 2025b] for the claim that larger synthetic datasets improve performance). These citations are not load-bearing: the taxonomy and qualitative analysis would stand without them, and SynQ's accuracy is reported as a surveyed result rather than used as a fitted input. No uniqueness theorem from the authors' prior work is invoked, no ansatz is smuggled in via citation, and no known result is renamed as new. The Table 1 comparability issues raised by the skeptic (ResNet-18 W4A4 rows alongside DeiT-Tiny W8A8 rows, heterogeneous image counts, and the apparent DeepInversion footnote mismatch) are legitimate internal-validity concerns for the claim of fair benchmarking, but they are not circularity: the table is an output compilation, not an input to a derivation. Consequently, there are no circular steps.
Assumptions & free parameters
assumptions (3)
- domain assumption Reported accuracies in the original papers are accurate and reproducible.
- ad hoc to paper The three-way taxonomy is exhaustive for the ZSQ literature.
- domain assumption The numbers in Table 1 are comparable across methods despite differing experimental protocols.
Cite this review
Pith. "Pith review of Zero-shot Quantization: A Comprehensive Survey." pith.science (2026). https://pith.science/paper/B7AUZ6KS
@misc{pith2026250509188,
author = {Pith},
title = {Pith review of: Zero-shot Quantization: A Comprehensive Survey},
year = {2026},
howpublished = {\url{https://pith.science/paper/B7AUZ6KS}},
note = {Machine review of arXiv:2505.09188}
}
read the original abstract
Network quantization has proven to be a powerful approach to reduce the memory and computational demands of deep learning models for deployment on resource-constrained devices. However, traditional quantization methods often rely on access to training data, which is impractical in many real-world scenarios due to privacy, security, or regulatory constraints. Zero-shot Quantization (ZSQ) emerges as a promising solution, achieving quantization without requiring any real data. In this paper, we provide a comprehensive overview of ZSQ methods and their recent advancements. First, we provide a formal definition of the ZSQ problem and highlight the key challenges. Then, we categorize the existing ZSQ methods into classes based on data generation strategies, and analyze their motivations, core ideas, and key takeaways. Lastly, we suggest future research directions to address the remaining limitations and advance the field of ZSQ. To the best of our knowledge, this paper is the first in-depth survey on ZSQ.
Figures
Reference graph
Works this paper leans on
-
[1]
Unified data-free compression: Pruning and quantization without fine-tuning
[Bai et al., 2023] Shipeng Bai, Jun Chen, Xintian Shen, Yix- uan Qian, and Yong Liu. Unified data-free compression: Pruning and quantization without fine-tuning. In ICCV,
work page 2023
-
[5]
Pea-kd: Parameter-efficient and accurate knowledge distillation on bert
[Cho and Kang, 2022] Ikhyun Cho and U Kang. Pea-kd: Parameter-efficient and accurate knowledge distillation on bert. PLOS ONE, 17(2):1–12, 02
work page 2022
-
[8]
[Deng et al., 2009] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR,
work page 2009
-
[10]
An image is worth 16x16 words: Trans- formers for image recognition at scale
[Dosovitskiy et al., 2020] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Min- derer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Trans- formers for image recognition at scale. In ICLR,
work page 2020
-
[11]
Sharpness-aware data generation for zero-shot quantization
[Dung et al., 2024] Hoang Anh Dung, Cuong Pham, Trung Le, Jianfei Cai, and Thanh-Toan Do. Sharpness-aware data generation for zero-shot quantization. In ICML,
work page 2024
-
[12]
Data-free quantization via pseudo-label fil- tering
[Fan et al., 2024] Chunxiao Fan, Ziqi Wang, Dan Guo, and Meng Wang. Data-free quantization via pseudo-label fil- tering. In CVPR,
work page 2024
-
[13]
A survey of quantization methods for efficient neural network inference
[Gholami et al., 2022] Amir Gholami, Sehoon Kim, Zhen Dong, Zhewei Yao, Michael W Mahoney, and Kurt Keutzer. A survey of quantization methods for efficient neural network inference. In Low-Power Computer Vision, pages 291–326. Chapman and Hall/CRC,
work page 2022
-
[14]
[Goodfellow et al., 2014] Ian Goodfellow, Jean Pouget- Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Gen- erative adversarial nets. In NeurIPS,
work page 2014
Show all 45 references
-
[16]
Deep learning with limited numerical precision
[Gupta et al., 2015] Suyog Gupta, Ankur Agrawal, Kailash Gopalakrishnan, and Pritish Narayanan. Deep learning with limited numerical precision. In ICML,
2015
-
[18]
Structured pruning for deep convolutional neural networks: A survey
[He and Xiao, 2024] Yang He and Lingao Xiao. Structured pruning for deep convolutional neural networks: A survey. IEEE Transactions on Pattern Analysis and Machine In- telligence, 46(5):2900–2919,
2024
-
[19]
Deep residual learning for image recog- nition
[He et al., 2016] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog- nition. In CVPR,
2016
-
[21]
Falcon: lightweight and accurate convo- lution based on depthwise separable convolution
[Jang et al., 2023] Jun-Gi Jang, Chun Quan, Hyun Dong Lee, and U Kang. Falcon: lightweight and accurate convo- lution based on depthwise separable convolution. Knowl. Inf. Syst., 65(5):2225–2249,
2023
-
[22]
[Kim et al., 2021] Junghun Kim, Jinhong Jung, and U. Kang. Compressing deep graph convolution network with multi- staged knowledge distillation. PLOS ONE, 16(8):1–18, 08
2021
-
[24]
Patch similarity aware data-free quantization for vision transformers
[Li et al., 2022] Zhikai Li, Liping Ma, Mengjuan Chen, Jun- rui Xiao, and Qingyi Gu. Patch similarity aware data-free quantization for vision transformers. In ECCV,
2022
-
[25]
Hard sample matters a lot in zero-shot quan- tization
[Li et al., 2023] Huantong Li, Xiangmiao Wu, Fanbing Lv, Daihai Liao, Thomas H Li, Yonggang Zhang, Bo Han, and Mingkui Tan. Hard sample matters a lot in zero-shot quan- tization. In CVPR,
2023
-
[26]
Zero-shot adversarial quantization
[Liu et al., 2021] Yuang Liu, Wei Zhang, and Jun Wang. Zero-shot adversarial quantization. In CVPR,
2021
-
[27]
Visual instruction tuning
[Liu et al., 2023] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS,
2023
-
[28]
Data-free quantization through weight equalization and bias correction
[Nagel et al., 2019] Markus Nagel, Mart van Baalen, Tijmen Blankevoort, and Max Welling. Data-free quantization through weight equalization and bias correction. In ICCV,
2019
-
[30]
A comprehensive survey of com- pression algorithms for language models
[Park et al., 2024b] Seungcheol Park, Jaehyeon Choi, Sojin Lee, and U Kang. A comprehensive survey of com- pression algorithms for language models. arXiv preprint arXiv:2401.15347,
-
[31]
Accurate sublayer pruning for large language models by exploiting latency and tunability information
[Park et al., 2025] Seungchul Park, Sojin Lee, Jongjin Kim, Jinsik Lee, Hyunsik Jo, and U Kang. Accurate sublayer pruning for large language models by exploiting latency and tunability information. In IJCAI,
2025
-
[32]
[Piao et al., 2022] Tairen Piao, Ikhyun Cho, and U. Kang. Sensimix: Sensitivity-aware 8-bit index & 1-bit value mixed precision quantization for bert compression. PLOS ONE, 17(4):1–22,
2022
-
[33]
Clamp-vit: contrastive data- free learning for adaptive post-training quantization of vits
[Ramachandran et al., 2024] Akshat Ramachandran, Souvik Kundu, and Tushar Krishna. Clamp-vit: contrastive data- free learning for adaptive post-training quantization of vits. In ECCV,
2024
-
[34]
Causal- dfq: Causality guided data-free network quantization
[Shang et al., 2023] Yuzhang Shang, Bingxin Xu, Gaowen Liu, Ramana Rao Kompella, and Yan Yan. Causal- dfq: Causality guided data-free network quantization. In CVPR,
2023
-
[35]
A generalized zero- shot quantization of deep convolutional neural networks via learned weights statistics
[Sharma et al., 2021] Prasen Kumar Sharma, Arun Abra- ham, and Vikram Nelvoy Rajendiran. A generalized zero- shot quantization of deep convolutional neural networks via learned weights statistics. IEEE Transactions on Mul- timedia, 25:953–965,
2021
-
[36]
Training data-efficient image transformers & distillation through attention
[Touvron et al., 2021] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablay- rolles, and Herv ´e J ´egou. Training data-efficient image transformers & distillation through attention. In ICML,
2021
-
[37]
Pruning has a disparate im- pact on model accuracy
[Tran et al., 2022] Cuong Tran, Ferdinando Fioretto, Jung- Eun Kim, and Rakshit Naidu. Pruning has a disparate im- pact on model accuracy. In NeurIPS,
2022
-
[38]
Qdrop: Randomly dropping quantization for extremely low-bit post-training quantization
[Wei et al., 2022] Xiuying Wei, Ruihao Gong, Yuhang Li, Xianglong Liu, and Fengwei Yu. Qdrop: Randomly dropping quantization for extremely low-bit post-training quantization. In ICLR,
2022
-
[39]
Towards a smaller stu- dent: Capacity dynamic distillation for efficient image re- trieval
[Xie et al., 2023] Yi Xie, Huaidong Zhang, Xuemiao Xu, Jianqing Zhu, and Shengfeng He. Towards a smaller stu- dent: Capacity dynamic distillation for efficient image re- trieval. In CVPR,
2023
-
[40]
Generative low-bitwidth data free quantization
[Xu et al., 2020] Shoukai Xu, Haokun Li, Bohan Zhuang, Jing Liu, Jiezhang Cao, Chuangrun Liang, and Mingkui Tan. Generative low-bitwidth data free quantization. In ECCV,
2020
-
[41]
Dreaming to distill: Data-free knowl- edge transfer via deepinversion
[Yin et al., 2020] Hongxu Yin, Pavlo Molchanov, Jose M Al- varez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowl- edge transfer via deepinversion. In CVPR,
2020
-
[42]
Knowledge extraction with no observable data
[Yoo et al., 2019] Jaemin Yoo, Minyong Cho, Taebum Kim, and U Kang. Knowledge extraction with no observable data. In Advances in Neural Information Processing Sys- tems,
2019
-
[43]
Diversifying sample generation for accurate data-free quantization
[Zhang et al., 2021] Xiangguo Zhang, Haotong Qin, Yifu Ding, Ruihao Gong, Qinghua Yan, Renshuai Tao, Yuhang Li, Fengwei Yu, and Xianglong Liu. Diversifying sample generation for accurate data-free quantization. In CVPR,
2021
-
[44]
Intraq: Learning synthetic images with intra-class heterogeneity for zero-shot network quantiza- tion
[Zhong et al., 2022] Yunshan Zhong, Mingbao Lin, Gongrui Nan, Jianzhuang Liu, Baochang Zhang, Yonghong Tian, and Rongrong Ji. Intraq: Learning synthetic images with intra-class heterogeneity for zero-shot network quantiza- tion. In CVPR,
2022
-
[45]
Autorecon: Neural ar- chitecture search-based reconstruction for data-free com- pression
[Zhu et al., 2021] Baozhou Zhu, Peter Hofstee, Johan Pel- tenburg, Jinho Lee, and Zaid Alars. Autorecon: Neural ar- chitecture search-based reconstruction for data-free com- pression. In IJCAI, 2021
2021
-
[2009]
Model compression and hardware ac- celeration for neural networks: A comprehensive survey
[Deng et al., 2020] Lei Deng, Guoqi Li, Song Han, Luping Shi, and Yuan Xie. Model compression and hardware ac- celeration for neural networks: A comprehensive survey. Proceedings of the IEEE, 108(4):485–532,
2020
-
[2014]
Squant: On-the-fly data-free quanti- zation via diagonal hessian approximation
[Guo et al., 2022] Cong Guo, Yuxian Qiu, Jingwen Leng, Xiaotian Gao, Chen Zhang, Yunxin Liu, Fan Yang, Yuhao Zhu, and Minyi Guo. Squant: On-the-fly data-free quanti- zation via diagonal hessian approximation. In ICLR,
2022
-
[2015]
The knowledge within: Meth- ods for data-free model compression
[Haroush et al., 2020] Matan Haroush, Itay Hubara, Elad Hoffer, and Daniel Soudry. The knowledge within: Meth- ods for data-free model compression. In CVPR,
2020
-
[2016]
Sparse model inversion: Efficient inversion of vision transformers for data-free applications
[Hu et al., 2024] Zixuan Hu, Yongxian Wei, Li Shen, Zhenyi Wang, Lei Li, Chun Yuan, and Dacheng Tao. Sparse model inversion: Efficient inversion of vision transformers for data-free applications. In ICML,
2024
-
[2019]
Up or down? adaptive rounding for post-training quanti- zation
[Nagel et al., 2020] Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. Up or down? adaptive rounding for post-training quanti- zation. In ICML,
2020
-
[2020]
Texq: zero- shot network quantization with texture feature distribution calibration
[Chen et al., 2023] Xinrui Chen, Yizhi Wang, Renao Yan, Yiqing Liu, Tian Guan, and Yonghong He. Texq: zero- shot network quantization with texture feature distribution calibration. In NeurIPS,
2023
-
[2021]
It’s all in the teacher: Zero-shot quantization brought closer to the teacher
[Choi et al., 2022] Kanghyun Choi, Hye Yoon Lee, Deokki Hong, Joonsang Yu, Noseong Park, Youngsok Kim, and Jinho Lee. It’s all in the teacher: Zero-shot quantization brought closer to the teacher. In CVPR,
2022
-
[2022]
Qimera: Data-free quantization with synthetic boundary supporting samples
[Choi et al., 2021] Kanghyun Choi, Deokki Hong, Noseong Park, Youngsok Kim, and Jinho Lee. Qimera: Data-free quantization with synthetic boundary supporting samples. In NeurIPS,
2021
-
[2023]
Robustness-guided image synthesis for data-free quanti- zation
[Bai et al., 2024] Jianhong Bai, Yuchen Yang, Huanpeng Chu, Hualiang Wang, Zuozhu Liu, Ruizhe Chen, Xi- aoxuan He, Lianrui Mu, Chengfei Cai, and Haoji Hu. Robustness-guided image synthesis for data-free quanti- zation. In AAAI,
2024
-
[2024]
Ze- roq: A novel zero shot quantization framework
[Cai et al., 2020] Yaohui Cai, Zhewei Yao, Zhen Dong, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. Ze- roq: A novel zero shot quantization framework. In CVPR,
2020
-
[2025]
Gra- dient harmonized single-stage detector
[Li et al., 2019] Buyu Li, Yu Liu, and Xiaogang Wang. Gra- dient harmonized single-stage detector. In AAAI,
2019
Reviewed August 15, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.