Pith. sign in

REVIEW 3 major objections 6 minor 39 references

Enhancing Generalization in Data-free Quantization via Mixup-class Prompting

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Mixup-class prompting: fusing two class labels in text prompts improves data-free quantization accuracy.

desk verdict A simple, reproducible prompt trick that reliably helps low-bit DFQ, with a theoretical story that overreaches. read the letter →

arxiv 2507.21947 v1 pith:T7XB53VH submitted 2025-07-29 cs.CV cs.AI

Jiwoong Park , Chaeun Lee , Yongseok Choi , Sein Park , Deokki Hong , Jungwook Choi This is my paper
classification cs.CVcs.AI
keywords post-trainingquantizationdata-freemixup-classprompttext-conditionedlatentdiffusionsyntheticcalibrationdatageneralizationgaplow-bitengineering
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Data-free post-training quantization (PTQ) needs synthetic images to calibrate a quantized model, but single-class prompts like 'a realistic photo of a kite' can generate images that do not match the intended class, because words like 'kite' are ambiguous. The paper proposes mixup-class prompting, which writes the prompt as '[template] [C1] and [C2]' to fuse two class labels at the text level. It claims this simple change increases the diversity and robustness of the synthetic calibration set, reduces the generalization gap as tracked by the empirical gradient norm of quantization parameters, and improves PTQ accuracy consistently across CNN and vision-transformer models. The strongest reported gains are in extremely low-bit settings, including new state-of-the-art accuracy for 2-bit weight, 4-bit activation (W2A4) quantization. A sympathetic reader would care because the method requires no architectural change, no extra hyperparameters, and no real calibration data.

What carries the argument

The machinery has three parts. First is the mixup-class prompt itself, defined as a text prompt '[template] [C1] and [C2]' with two randomly sampled class labels, which is fed to a text-conditioned latent diffusion model to synthesize calibration images. Second is a generalization-gap criterion: the paper adopts an upper bound from noisy-gradient non-convex optimization that ties the generalization gap to the empirical gradient norm of the optimized parameters, and it restricts those parameters to the quantization parameters of PTQ; the trace of gradient norms over calibration steps is used to rank candidate calibration sets. Third is the RPC-FID metric, a per-class Fréchet-Inception-Distance ratio that quantifies how far a synthetic class distribution sits from real training data relative to natural intra-class variation.

What would settle it

Run the same PTQ algorithm on many prompt strategies and record both the gradient-norm trace and the resulting quantized accuracy on a common test set; if a strategy with consistently lower gradient norms yields worse accuracy than a higher-norm strategy, the claimed link between gradient norm and generalization for PTQ calibration is not capturing what matters.

Watch

Extended reading notes

Core claim

The paper's central discovery is that where you mix matters: instead of mixing images after generation (CutMix, Mixup, ResizeMix), mixing class labels inside the text prompt before diffusion generates calibration images with lower gradient norms for quantization parameters and better downstream quantized accuracy. Concretely, with Genie-M for CNNs and RePQ-ViT for ViTs, prompts of the form $[\text{template}] [C_1] \text{ and } [C_2]$ outperform single-class prompting, established prompt-engineering variants, and pixel-level augmentations across W4A4, W3A3, W2A4, W5A5, and W8A8 settings. The paper also reports that random class pairing beats semantic-similarity-based pairing, and that two classes per prompt work better than three or four. Its analytical claim is that the mixup-class prompt lowers the upper bound on the generalization gap from Eq. (3), making calibration optimization more stable.

Load-bearing premise

The argument's load-bearing premise is that the theoretical generalization-gap bound derived for noisy stochastic gradient training of full model parameters still applies when only quantization parameters are optimized blockwise on a synthetic calibration set; Section 3.1 asserts this restriction 'without loss of generality' but does not prove the bound's assumptions hold.

Editorial extensions

If this is right

  • Calibration sets generated with mixup-class prompts should yield more accurate quantized models than single-class-prompt sets at the same budget, especially at W2A4 and other low-bit settings.
  • Because the prompt change is architecture-agnostic, the same prompting trick can be dropped into any existing DFQ pipeline that uses a text-conditioned diffusion model.
  • Random pairing of classes is a feature, not a bug: forcing semantic similarity between the two mixed classes does not improve accuracy, so no class-embedding selection step is needed.
  • The gradient-norm trace provides a practical, training-free indicator for judging whether a candidate calibration set will generalize before running full quantization.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • My inference: the gradient-norm trace could be repurposed as a prompt-selection or filtering criterion for any LDM-based synthetic calibration pipeline, potentially replacing extra synthetic-image filtering and its compute overhead.
  • My inference: because text-level mixing is orthogonal to pixel-level augmentation, combining mixup-class prompts with image-space mixup or augmentation on generated images may yield further gains; the paper does not test this combination.
  • My inference: the 'two classes are best' result suggests a sweet spot between prompt clarity and label diversity, but the proposed mechanism (polysemy dilution) would predict that classes with high polysemy benefit most; a targeted study on high-RPC-FID classes alone would test that prediction.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes mixup-class prompt, a text-prompting strategy for data-free post-training quantization (DFQ) that fuses two class labels ("[template] [C1] and [C2]") when generating synthetic calibration images with a latent diffusion model. The authors report that this simple change to the prompt reduces the empirical gradient norm with respect to quantization parameters, which they connect via an imported generalization-gap bound to improved generalization of the quantized model. They evaluate the method on CNNs (ResNet20/18/50, MobileNetV2) and ViTs (ViT, DeiT, Swin) under W4A4, W3A3, W2A4, W8A8, and W5A5 settings, comparing with GenQ, SADAG, MimiQ, and other DFQ baselines, and report state-of-the-art or competitive accuracy, especially at low bit-widths such as W2A4 on ImageNet. The paper also introduces RPC-FID as a diagnostic for polysemy-induced out-of-distribution synthetic images and includes ablations on augmentation methods, prompt engineering variants, other diffusion models, semantic similarity of class pairs, and the number of classes in the prompt.

Significance. If the empirical claims hold, the method is practically valuable: it is a simple, training-free change to the text prompt that requires no architectural modification and no additional generation cost beyond standard single-class prompting, and it is shown to work across several CNN and ViT architectures, multiple bit-widths, and three different Stable Diffusion variants. The release of code is a further strength. The paper also provides a concrete diagnostic (RPC-FID) for the polysemy problem, which the community may find useful. However, the paper's central analytical contribution—the claimed theoretical link between gradient norms and generalization in PTQ via Eq. (3)—is not established for the actual PTQ optimizers used, and some of the empirical gains are marginal or reversed in specific configurations. The contribution is therefore best viewed as an empirical finding with a plausible but unproven mechanism, and the presentation should be adjusted accordingly.

major comments (3)
  1. [Section 3.1, Eq. (3)] The generalization-gap bound is imported from stochastic-gradient non-convex optimization theory (An et al. [1]; Li et al. [16]) and the paper states that restricting θ to quantization parameters is "without loss of generality." This is not established: the bound assumes mini-batch SGD updates with Gaussian gradient noise and a learning-rate schedule γ_t, whereas Genie-M and RePQ-ViT are blockwise/scale-reparameterization optimizers that update only quantization parameters with different update rules. No argument is given that the bound's assumptions (e.g., Gaussian noise, full-parameter updates) transfer to these algorithms. Therefore the theoretical conclusion that a smaller cumulative gradient norm provably implies a smaller generalization gap for the actual PTQ procedure is unsupported; the analysis in Section 4 remains a plausible analogy rather than a quantitative explanation.
  2. [Section 5, Tables 1 and 2] The claim that the method "consistently outperforms" prior DFQ methods is stronger than the data. Several differences are within a few hundredths of a percent (CIFAR-100 W3A3 ResNet20: 65.96 vs 65.94; W4A4 ResNet20: 69.11 vs 69.11; ImageNet W2A4 ResNet18: 65.90 vs 65.72). In Table 2, ViT-B W8A8 is below MimiQ (84.32 vs 85.17) and below the real-data result (84.41), and Swin-B W4A4 leads MimiQ by only 0.04% (73.53 vs 73.49). Since no multiple seeds or error bars are reported, these marginal differences are within plausible run-to-run variance. Please report multiple seeds and temper the "consistently outperforms" claim, or restrict it to the low-bit configurations where the gains are substantial.
  3. [Section 4 and Fig. 3 vs Table 1] The proposed criterion that a lower gradient norm with respect to quantization parameters implies a smaller generalization gap is in tension with the accuracy results. Fig. 3 shows that mixup-class prompt achieves a lower gradient norm than real images, yet Table 1 shows that real-data calibration yields higher accuracy than mixup-class in nearly every row (e.g., ImageNet ResNet50 W3A3: Real 72.85 vs Ours 72.76; W2A4: Real 70.88 vs Ours 70.78). If the gradient-norm bound were the right predictor, mixup-class should match or exceed real-data accuracy. The paper needs to explain this discrepancy, or the gradient-norm analysis should be presented as descriptive rather than as the mechanism for the accuracy improvements.
minor comments (6)
  1. [Section 3.1, Eq. (1)] In the definition of RPC-FID, the notation is inconsistent: m_i and C_i are defined as moments of X_i, but the surrounding text says they are moments of feature vectors f(X_i). Please align the notation.
  2. [Author affiliations] The footnote markers for the email addresses are inconsistent; for example, "1yongseokchoi@sk.com" should use affiliation marker 3 rather than 1.
  3. [Section 3.2 vs Section 6.5] The paper states in Section 3.2 that mixup-class prompt "eliminates the need for additional hyperparameter tuning," but the number of classes per prompt is itself a hyperparameter, and Section 6.5 shows that choosing 2 classes is important for accuracy. Please qualify this claim.
  4. [Section 5.1] The text says CNN models are quantized with "Genie-M", but the baseline table lists "Genie"; please clarify whether these refer to the same algorithm or to different variants.
  5. [Table 6] The "High" and "Low" similarity strategies give identical accuracy for ResNet18 (65.71); this looks suspicious and should be checked or reported with uncertainty.
  6. [Figure 3 caption] The phrase "the trace of gradient norms" should be "the traces of gradient norms" or "gradient-norm traces" for grammatical clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; empirical gains are benchmark-driven and the analytical bound is imported from external SGD theory, not derived from the method's own definitions.

full rationale

The central claim (that mixup-class prompting improves PTQ) is tested against external baselines (GenQ, MimiQ, etc.) in Tables 1-5. The prompt format is not defined in terms of the reported accuracy, and no fitted parameter is presented as a prediction. The gradient-norm observations (Fig. 3) and RPC-FID diagnostics are measurements, not premises from which the accuracy results are derived. The generalization-gap bound (Eq. 3) comes from prior SGD theory by external authors (An et al., Li et al.); its transfer to blockwise PTQ optimizers is asserted with 'without loss of generality' but is not shown to hold, which is a validity concern rather than a circular reduction. The choice of two classes is an explicit ablation (Table 7). The only author-overlapping citation (Qimera, by Deokki Hong) is a baseline, not load-bearing. No equation reduces to another by construction, and no result is forced by self-citation, so the paper is not circular.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the LDM's ability to compose two class labels into useful calibration images and on the transfer of a generalization-gap bound from SGD theory to PTQ. The number of mixed classes (2) was chosen after ablation on the test datasets. No new physical entities or fitted constants are introduced.

free parameters (1)
  • Number of classes per mixup prompt = 2
    Chosen after an ablation on ImageNet test data (Table 7 compares 2, 3, and 4 classes). The final method uses 2 because it gave the highest accuracy; this is a hand-chosen design decision made on the same evaluation sets used for the headline results.
assumptions (3)
  • domain assumption The generalization-gap bound in Eq. (3), proven for noisy-gradient methods in non-convex optimization, applies to PTQ optimization restricted to quantization parameters.
    Used in Section 3.1 to justify treating the gradient norm of quantization parameters as a proxy for generalization. No proof is given that the theorem's assumptions hold for blockwise PTQ optimizers such as Genie-M and RePQ-ViT.
  • domain assumption A text-conditioned LDM (Stable Diffusion) can compose two class labels in one prompt into images that are useful calibration data.
    This is the core mechanism of mixup-class prompt. It relies on the generative model's compositional abilities, which are plausible but not guaranteed for all class pairs, especially fine-grained or rare classes.
  • domain assumption The newly proposed RPC-FID metric measures polysemy-induced distribution shift between synthetic and real data.
    Introduced in Section 3.1, Eq. (1). The denominator uses FID between two disjoint subsets of real training data, which may be unstable for small samples, and the metric is not validated against accuracy before being used as motivation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Generalization in Data-free Quantization via Mixup-class Prompting." pith.science (2026). https://pith.science/paper/T7XB53VH

@misc{pith2026250721947,
  author       = {Pith},
  title        = {Pith review of: Enhancing Generalization in Data-free Quantization via Mixup-class Prompting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T7XB53VH}},
  note         = {Machine review of arXiv:2507.21947}
}
read the original abstract

Post-training quantization (PTQ) improves efficiency but struggles with limited calibration data, especially under privacy constraints. Data-free quantization (DFQ) mitigates this by generating synthetic images using generative models such as generative adversarial networks (GANs) and text-conditioned latent diffusion models (LDMs), while applying existing PTQ algorithms. However, the relationship between generated synthetic images and the generalizability of the quantized model during PTQ remains underexplored. Without investigating this relationship, synthetic images generated by previous prompt engineering methods based on single-class prompts suffer from issues such as polysemy, leading to performance degradation. We propose \textbf{mixup-class prompt}, a mixup-based text prompting strategy that fuses multiple class labels at the text prompt level to generate diverse, robust synthetic data. This approach enhances generalization, and improves optimization stability in PTQ. We provide quantitative insights through gradient norm and generalization error analysis. Experiments on convolutional neural networks (CNNs) and vision transformers (ViTs) show that our method consistently outperforms state-of-the-art DFQ methods like GenQ. Furthermore, it pushes the performance boundary in extremely low-bit scenarios, achieving new state-of-the-art accuracy in challenging 2-bit weight, 4-bit activation (W2A4) quantization.

Figures

Figures reproduced from arXiv: 2507.21947 by the authors.

Figure 1
Figure 1. Top 50% classes of ImageNet sorted by RPC-FID. A [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. (a) The overview of pipeline for our method and (b) examples of generated images using [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. t-SNE plots for the feature map of ResNet50. Samples [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (1 more)
Figure 3
Figure 3. Figure 3: Comparison of the trace of gradient norms with respect [PITH_FULL_IMAGE:figures/full_fig_p005_3.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 31 canonical work pages

  1. [1]

    Can we use gradient norm as a measure of generalization error for model selection in prac- tice? 2021

    Haozhe An, Haoyi Xiong, Xuhong Li, Xingjian Li, Dejing Dou, and Zhanxing Zhu. Can we use gradient norm as a measure of generalization error for model selection in prac- tice? 2021. 3 8

    work page 2021

  2. [16]

    On generaliza- tion error bounds of noisy gradient methods for non-convex learning

    Jian Li, Xuanyuan Luo, and Mingda Qiao. On generaliza- tion error bounds of noisy gradient methods for non-convex learning. arXiv preprint, 2019. 3

    work page 2019

  3. [2]

    Qimera: Data-free quantization with synthetic boundary supporting samples

    Kanghyun Choi, Deokki Hong, Noseong Park, Youngsok Kim, and Jinho Lee. Qimera: Data-free quantization with synthetic boundary supporting samples. NeurIPS, 34, 2021. 6

    work page 2021

  4. [3]

    MimiQ: Low-Bit Data-Free Quantization of Vision Transformers with Encouraging Inter-Head Attention Similarity

    Kanghyun Choi, Hye Yoon Lee, Dain Kwon, SunJong Park, Kyuyeun Kim, Noseong Park, and Jinho Lee. Mimiq: Low- bit data-free quantization of vision transformers with en- couraging inter-head attention similarity. arXiv preprint arXiv:2407.20021, 2024. 6

    work page Pith review arXiv 2024

  5. [4]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6

    arXiv 2010

  6. [5]

    Sharpness-aware data generation for zero- shot quantization

    Hoang Anh Dung, Cuong Pham, Trung Le, Jianfei Cai, and Thanh-Toan Do. Sharpness-aware data generation for zero- shot quantization. In ICML, 2024. 6

    work page 2024

  7. [6]

    Scaling recti- fied flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling recti- fied flow transformers for high-resolution image synthesis. In Forty-first international conference on machine learning,

  8. [7]

    Scaling laws of synthetic images for model training

    Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In CVPR, pages 7382– 7392, 2024. 3

    work page 2024

Show all 39 references
  1. [8]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770–778, 2016. 6

    2016

  2. [9]

    Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022

    Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022. 3, 7

    2022 arXiv

  3. [10]

    Gans trained by a two time-scale update rule converge to a local nash equilib- rium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium. Advances in neural information processing systems , 30, 2017. 3

    2017

  4. [11]

    Difficulty diversity and plausibility: Dynamic data-free quantization

    Cheeun Hong, Sungyong Baik, Junghun Oh, and Ky- oung Mu Lee. Difficulty diversity and plausibility: Dynamic data-free quantization. In Proceedings of the Winter Con- ference on Applications of Computer Vision (WACV), pages 7531–7540, 2025. 8

    2025

  5. [12]

    Diffusemix: Label- preserving data augmentation with diffusion models

    Khawar Islam, Muhammad Zaigham Zaheer, Arif Mah- mood, and Karthik Nandakumar. Diffusemix: Label- preserving data augmentation with diffusion models. In CVPR, 2024. 3

    2024

  6. [13]

    Ge- nie: show me the data for quantization

    Yongkweon Jeon, Chungman Lee, and Ho-young Kim. Ge- nie: show me the data for quantization. In CVPR, 2023. 5, 6

    2023

  7. [14]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky et al. Learning multiple layers of features from tiny images. 2009. 6

    2009

  8. [15]

    Flexround: Learnable rounding based on element- wise division for post-training quantization

    Jung Hyun Lee, Jeonghoon Kim, Se Jung Kwon, and Dong- soo Lee. Flexround: Learnable rounding based on element- wise division for post-training quantization. In Interna- tional Conference on Machine Learning , pages 18913– 18939. PMLR, 2023. 2

    2023

  9. [17]

    Brecq: Pushing the limit of post-training quantization by block reconstruc- tion

    Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. Brecq: Pushing the limit of post-training quantization by block reconstruc- tion. arXiv preprint, 2021. 2

    2021

  10. [18]

    Mixmix: All you need for data-free compression are feature and data mix- ing

    Yuhang Li, Feng Zhu, Ruihao Gong, Mingzhu Shen, Xin Dong, Fengwei Yu, Shaoqing Lu, and Shi Gu. Mixmix: All you need for data-free compression are feature and data mix- ing. In ICCV, pages 4410–4419, 2021. 2

    2021

  11. [19]

    Genq: Quantization in low data regimes with generative synthetic data

    Yuhang Li, Youngeun Kim, Donghyun Lee, Souvik Kundu, and Priyadarshini Panda. Genq: Quantization in low data regimes with generative synthetic data. InECCV, pages 216–

  12. [20]

    Patch similarity aware data-free quantization for vision transformers

    Zhikai Li, Liping Ma, Mengjuan Chen, Junrui Xiao, and Qingyi Gu. Patch similarity aware data-free quantization for vision transformers. In ECCV. Springer, 2022. 6

    2022

  13. [21]

    Repq- vit: Scale reparameterization for post-training quantization of vision transformers

    Zhikai Li, Junrui Xiao, Lianwei Yang, and Qingyi Gu. Repq- vit: Scale reparameterization for post-training quantization of vision transformers. In ICCV, 2023. 6

    2023

  14. [22]

    Pd-quant: Post-training quantiza- tion based on prediction difference metric

    Jiawei Liu, Lin Niu, Zhihang Yuan, Dawei Yang, Xinggang Wang, and Wenyu Liu. Pd-quant: Post-training quantiza- tion based on prediction difference metric. In CVPR, pages 24427–24437, 2023. 2

    2023

  15. [23]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 6

    2021

  16. [24]

    George A. Miller. Wordnet: a lexical database for english. Commun. ACM, 1995. 7

    1995

  17. [25]

    Metaaug: Meta-data augmentation for post-training quanti- zation

    Cuong Pham, Anh Dung Hoang, Cuong C Nguyen, Trung Le, Dinh Phung, Gustavo Carneiro, and Thanh-Toan Do. Metaaug: Meta-data augmentation for post-training quanti- zation. In ECCV. Springer, 2024. 1, 2, 3

    2024

  18. [26]

    Adaptive data-free quantization

    Biao Qian, Yang Wang, Richang Hong, and Meng Wang. Adaptive data-free quantization. In CVPR, 2023. 6

    2023

  19. [27]

    Resizemix: Mixing data with preserved object information and true labels

    Jie Qin, Jiemin Fang, Qian Zhang, Wenyu Liu, Xingang Wang, and Xinggang Wang. Resizemix: Mixing data with preserved object information and true labels. arXiv preprint arXiv:2012.11101, 2020. 2, 5

    2012 arXiv

  20. [28]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. In CVPR, 2022. 1, 3, 7

    2022

  21. [29]

    Imagenet large scale visual recognition challenge

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, San- jeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of com- puter vision, 115:211–252, 2015. 3, 6

    2015

  22. [30]

    Mobilenetv2: Inverted residuals and linear bottlenecks

    Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh- moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 6

    2018

  23. [31]

    Fake it till you make it: Learning trans- ferable representations from synthetic imagenet clones

    Mert B ¨ulent Sarıyıldız, Karteek Alahari, Diane Larlus, and Yannis Kalantidis. Fake it till you make it: Learning trans- ferable representations from synthetic imagenet clones. In CVPR, 2023. 3, 4, 7 9

    2023

  24. [32]

    Learning vision from models rivals learning vision from data

    Yonglong Tian, Lijie Fan, Kaifeng Chen, Dina Katabi, Dilip Krishnan, and Phillip Isola. Learning vision from models rivals learning vision from data. In CVPR, 2024. 3

    2024

  25. [33]

    Training data-efficient image transformers & distillation through at- tention

    Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv ´e J´egou. Training data-efficient image transformers & distillation through at- tention. In ICML. PMLR, 2021. 6

    2021

  26. [34]

    Enhance im- age classification via inter-class image mixup with diffusion model

    Zhicai Wang, Longhui Wei, Tan Wang, Heyu Chen, Yanbin Hao, Xiang Wang, Xiangnan He, and Qi Tian. Enhance im- age classification via inter-class image mixup with diffusion model. In CVPR, 2024. 3

    2024

  27. [35]

    Qdrop: Randomly dropping quantization for extremely low-bit post-training quantization

    Xiuying Wei, Ruihao Gong, Yuhang Li, Xianglong Liu, and Fengwei Yu. Qdrop: Randomly dropping quantization for extremely low-bit post-training quantization. arXiv preprint arXiv:2203.05740, 2022. 1, 2, 3

    2022 arXiv

  28. [36]

    Schr \”{o} dinger’s bat: Diffusion models sometimes generate polysemous words in superposition

    Jennifer C White and Ryan Cotterell. Schr \”{o} dinger’s bat: Diffusion models sometimes generate polysemous words in superposition. arXiv preprint arXiv:2211.13095 ,

    arXiv

  29. [37]

    Cutmix: Regu- larization strategy to train strong classifiers with localizable features

    Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regu- larization strategy to train strong classifiers with localizable features. In ICCV, 2019. 2

    2019

  30. [38]

    Dauphin, and David Lopez-Paz

    Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimiza- tion, 2018. 2, 4, 5 10

    2018

  31. [235]

    1, 2, 3, 4, 6

    Springer, 2025. 1, 2, 3, 4, 6

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.