Pith. sign in

REVIEW 4 major objections 7 minor 31 references

PQCAD-DM: Progressive Quantization and Calibration-Assisted Distillation for Extremely Efficient Diffusion Model

T0 review · 4 major / 7 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 PQCAD-DM halves the sampling steps of a quantized diffusion model while keeping or improving image quality.

desk verdict Plausible compression recipe with one genuinely interesting idea, but the central CAD loss is undefined as written and the reported FID gains are mostly within noise; worth reviewing after heavy revision, not citable yet. read the letter →

arxiv 2506.16776 v1 pith:GR6YBPH3 submitted 2025-06-20 cs.CV cs.AI

Beomseok Ko , Hyeryung Jang This is my paper
classification cs.CVcs.AI
keywords diffusionmodelsmodelcompressionpost-trainingquantizationknowledgedistillationprogressivecalibration-assistedEarth-Moverdistanceimagegeneration
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

Diffusion models generate high-quality images but pay for it with hundreds of sampling steps and heavy compute. PQCAD-DM is a compression pipeline that attacks both costs at once: it first quantizes a pretrained diffusion model down to low bit-widths in two progressive stages, then distills that quantized model into a student that needs only half the sampling steps. The central claim is that this combination does not degrade generative quality: at 50 steps and 4/8-bit weights/activations, PQCAD-DM matches or beats the 100-step full-precision baselines on FID, sFID, and Inception Score across CIFAR-10, ImageNet, CelebA-HQ, LSUN-Bedrooms, and LSUN-Churches. The practical promise is that image-generation models become small and fast enough for edge devices without retraining from scratch.

What carries the argument

The machinery has three parts. Progressive Quantization (PQ) first quantizes weights to an intermediate bit-width $\tau$ (typically double the target $\kappa$), refining block-wise outputs against a time-aware calibration set, and switches to $\kappa$ only when a momentum-based detector sees the perturbation-loss gradient flatten (window size 500, threshold 0.04). Calibration-Assisted Distillation (CAD) then trains a student from the quantized teacher with two losses: the standard distillation loss $w(\lambda_t)\|\tilde{x} - \hat{x}_\theta(z_t)\|_2^2$ against the teacher's two-step target $\tilde{x}$, plus a calibration term $\lambda \cdot \text{EM}(C_{DC}, \hat{x}_\theta(z_t))$ that uses Earth-Mover distance to align the student's output to full-precision calibration data $C_{DC}$, collected from the FP model at even time steps with time-conditioned uniform sampling. The third part is the dual calibration dataset design: $C_{QC}$ for quantization and $C_{DC}$ for distillation.

What would settle it

Measure the signal-to-noise ratio (SNR) of the student's input at student step k and the full-precision model's input at teacher step 2k as defined by the noise schedule; if these SNRs differ by more than the inter-step variation of the teacher model, the time-correspondence heuristic is broken and the calibration loss is misaligned.

Watch

Extended reading notes

Core claim

The paper's core claim is that progressive quantization and calibration-assisted distillation can be combined so that the two errors do not add up: the quantized teacher's mistakes, which would normally be amplified when distilling a student, are corrected by a calibration loss that compares the student's output to the full-precision model's output on a matching time step. Concretely, the authors show that PQCAD-DM, built on top of PTQ4DM, Q-Diff, or TFMQ-DM, reaches the same or better FID/sFID/IS at 50 sampling steps than those baselines do at 100 steps, and on CelebA-HQ with TFMQ-DM the 50-step quantized model (FID 8.57) actually beats the full-precision 100-step model (FID 8.74). At 4/8-bit weight/activation precision, the method also reduces geometric-mean weight perturbation by roughly 30 percent relative to its base quantizer, which is the mechanism the paper identifies for the quality retention.

Load-bearing premise

The whole gain depends on the assumption that the full-precision model's outputs at even time steps are correct regression targets for the quantized student's outputs at half those steps; if the two models are not aligned in noise level, the calibration loss teaches the wrong thing.

Editorial extensions

If this is right

  • If the claim holds, any existing PTQ method for diffusion models (PTQ4DM, Q-Diff, TFMQ-DM) can be upgraded to halve sampling steps without sacrificing FID/sFID, simply by plugging in PQ and CAD.
  • At 4/8-bit weight/activation precision, PQCAD-DM cuts inference time roughly in half and halves bit-operations compared to the base quantizer at full steps, directly enabling larger batch sizes and higher throughput on memory-limited GPUs.
  • The progressive-quantization stage alone improves fixed-bit baselines even without distillation, e.g., on CIFAR-10 PTQ4DM+PQ drops FID from 19.59 to 13.83, showing that lower-precision quality loss is reversible with better transition timing.
  • Distillation with calibration beats plain distillation on quantized teachers: on LSUN-Churches Q-Diff, CAD improves FID from 5.98 to 4.54, and on LSUN-Bedrooms TFMQ-DM, from 4.97 to 3.10, demonstrating that full-precision calibration data compensates for teacher inaccuracy.

Reading between the lines

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

  • The paper fixes the student-teacher time correspondence as $t_{\text{student}} = t_{\text{teacher}}/2$; an adaptive step-alignment scheme, which the authors do not explore, might extend the method to other step-reduction ratios or to non-uniform noise schedules.
  • The momentum-based transition detector is presented inside a diffusion-model pipeline, but it is really a generic schedule for when to drop precision; it could plausibly benefit quantization of other temporally varying models (e.g., video or audio generators), which the paper does not test.
  • One could read the EM calibration loss as a distribution-matching regularizer; a natural testable extension is whether the same term helps when the teacher and student use different samplers (e.g., DDIM vs a higher-order solver), which is not benchmarked here.
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

4 major / 7 minor

Summary. The paper proposes PQCAD-DM, a hybrid compression framework for diffusion models that combines Progressive Quantization (PQ) and Calibration-Assisted Distillation (CAD). PQ quantizes weights and activations in two stages (first to an intermediate bit-width τ, then to the target bit-width κ) with a momentum-based bit-transition detector, while CAD distills the quantized teacher into a student that uses half the sampling steps, guided by a calibration set generated by the full-precision model. The paper reports FID, sFID, and IS results on CIFAR-10, ImageNet, CelebA-HQ, and LSUN-Bedrooms/Churches, and claims that the method halves inference time while maintaining or improving generative quality relative to fixed-bit post-training quantization baselines, even at 4/8-bit weights/activations.

Significance. The engineering idea of combining progressive quantization with distillation for diffusion models is timely, and the experimental sweep across datasets and bit-widths is broad. The paper's strengths include explicit algorithms for calibration-data collection and quantization, a momentum-based transition mechanism that is clearly stated, and an ablation comparing EM distance with other distributional distances in the calibration loss. However, the central claim cannot currently be verified from the manuscript: Eq. (8) defines the calibration loss in a way that is inconsistent with Algorithm 4, and the reported quality gains are mostly within typical FID noise with no error bars, seeds, or uncertainty quantification. If the CAD loss can be unambiguously redefined and the experiments repeated with proper statistical support, the method may be a useful contribution; as written, the evidence is not conclusive.

major comments (4)
  1. [Section IV-D, Eq. (8), Algorithms 1 and 4] The calibration loss LossCD is not well-defined as written. Eq. (8) defines it as λ·EM(CDC, x̂θκ(zt)), where CDC is a set of full-precision model outputs collected only at even time steps (Algorithm 1, line 8), while x̂θκ(zt) is a single student output. The manuscript never defines the Earth-Mover distance between a set and a single point, and Algorithm 4 line 12 instead computes λ·EM(xt, x̂θκ(zt)) with a single sampled calibration image xt. Moreover, Algorithm 4 line 6 samples t = i/T with i uniform on [1,T], so for every odd i the set CDC has no entry at that time step; the text's proposed 'time-conditioned uniform sampling' mapping (student step t/2 to teacher step t) is never implemented as an even-step restriction or a specific pairing rule. Because the central claim of maintaining quality at 50 steps rests on this loss, the experiments cannot be reproduced from the paper as it stands.
  2. [Section V-A, Section V-B, and Table I] The reported FID, sFID, and IS improvements over the fixed-bit baselines are mostly 0.04 to 0.32, with no standard deviations, error bars, or number of seeds. For example, Table I shows PQCAD-DM(TFMQ-DM) at 8/8 bits improving on LSUN-Bedrooms by only 0.04 FID (3.14 to 3.10) and on LSUN-Churches by 0.09 (4.01 to 3.92). These deltas are small relative to run-to-run variation typically observed for these metrics, so the claim that PQCAD-DM 'maintains or surpasses' baseline performance is not statistically supported. The abstract's 'improving FID by 0.17 on average' is likewise within this noise band. The authors should provide multiple seeds, error bars, or a significance test.
  3. [Section V-B and Table II] The efficiency claims are overstated and numerically inconsistent. The text says PQCAD-DM 'halves inference time' and 'runs 1.6× faster' than Q-Diff, but Table II shows 204.65 vs. 340.11 minutes on LSUN-Bedroom (a 1.66× speedup) and 17.44 vs. 30.60 minutes on CIFAR-10 (a 1.75× speedup), not a halving. Additionally, the statement 'double the efficiency in terms of GBOPs (3,240 vs. 6,480)' is numerically a halving of bit-operations, not a doubling of efficiency; throughput is roughly doubled, but the wording should be corrected to say that GBOPs are halved.
  4. [Section IV-B and Section IV-D] The time-step correspondence assumption underlying CAD is heuristic and unvalidated. The method assumes that full-precision outputs at teacher time step t are valid regression targets for the student at time step t/2, with only the phrase 'time-conditioned uniform sampling' as motivation. If the noise levels of zt differ between the teacher and student trajectories, the EM calibration loss injects mismatched supervision, and the paper provides no analysis or sensitivity study for this mapping. Since this is a free design choice central to the method's novelty, it should be investigated explicitly.
minor comments (7)
  1. [Fig. 5] In the bottom table of Fig. 5, the ΔPert values are printed as '∆311', '∆341', '∆367', etc., which appear to be missing decimal points (likely 3.11, 3.41, 3.67). Please clarify the notation.
  2. [Section V-B, Table I] The text states 'On LSUN-Churches, it achieves an FID of 3.64 and sFID of 7.46', but Table I in the 4/8-bit row for PQCAD-DM(TFMQ-DM) lists those values under LSUN-Bedrooms; the LSUN-Churches entries are 3.82 and 11.44. The dataset label is misassigned.
  3. [Section VI] The 'Discussion and Future Work' section is empty; the section header is immediately followed by the Conclusion. Please either provide content or remove the header.
  4. [Algorithm 1, line 9] There is a typo: 'Update CDC ← C DC ∪' should read 'Update CDC ← CDC ∪'.
  5. [Section II and Section III] Several citations are left as '[ ?]' in the related-work discussion of efficient diffusion models and post-training quantization; these unresolved references should be completed.
  6. [Table II] The dataset names are inconsistent: 'LSUN-Bedroom' appears in Table II while 'LSUN-Bedrooms' is used in Table I and the text; please use one consistent name.
  7. [Fig. 1] The right panel of Fig. 1 is referenced in the text, but the MSE comparison it depicts is not described; please add a sentence explaining the experimental setup and the takeaway.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: PQCAD-DM's central claims are evaluated against external FID/IS benchmarks, and its distillation targets are standard teacher/FP outputs rather than redefinitions of the reported metrics.

full rationale

The derivation chain of PQCAD-DM is not circular. The progressive-quantization loss (Eq. 6) minimizes block-wise reconstruction error relative to the full-precision model, and the proposed bit-transition criterion is a heuristic on perturbation loss; these are training objectives, not predictions. The calibration-assisted distillation loss (Eq. 8) combines the usual progressive-distillation target x̃ (the two-step DDIM output of the quantized teacher) with an Earth-Mover term against a calibration set CDC generated by the full-precision model. This is standard self-referential supervision in knowledge distillation: the student is trained to imitate the FP/teacher model, which is exactly the intended mechanism, not a hidden equivalencing of input and claim. The paper's headline results are FID, sFID, and IS computed on 50K generated images against real data distributions (Table I), so the central 'maintains generative quality' claim is externally falsifiable and does not reduce to the calibration objective by construction. There are no load-bearing self-citations, no imported uniqueness theorems, and no fitted constant is presented as a prediction. The skeptic's concern about Eq. 8's EM distance being underspecified (a set vs. a single sample, odd-time-step sampling in Alg. 4) is a reproducibility/correctness issue, not a circularity issue, and therefore does not affect this score.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The central claim rests on several empirical assumptions about diffusion activation statistics, the validity of a Hessian-based quantization objective, and the alignment of FP calibration outputs with student time steps. None of these are formalized beyond the paper's own heuristics, and the hyperparameters that control them are not swept.

free parameters (7)
  • CAD loss weight lambda = not reported
    Hyperparameter balancing LossKD and LossCD in Eq. (8); no value or sensitivity analysis is given in the experiments.
  • Momentum coefficient beta = 0.9
    Used in momentum-based bit transition (Alg. 3); fixed by hand, no sensitivity study.
  • Transition threshold pi = 0.04
    Bit-transition criterion in Alg. 3; chosen by hand and could shift the PQ/kappa trade-off.
  • Loss window size W = 500
    Fixed-size window for perturbation loss averaging in Alg. 3; no sensitivity analysis is reported.
  • PQ learning rate gamma = not reported
    Gradient update in Eq. (6) and Alg. 2; value absent from the experimental setup.
  • Calibration sampling counts n_t, n_fixed, n_max = not reported
    Sizes of CQC and CDC in Alg. 1 are not specified, affecting quantization calibration quality and distillation signal.
  • Intermediate bit-width tau = tau = 2 kappa
    Choice of double target bit-width is a design decision; no ablation with other ratios is provided.
assumptions (4)
  • domain assumption Taylor expansion truncation of the quantization objective (Eq. 4-5) is accurate at low bit-widths
    The PQ gradient updates minimize a Hessian-weighted perturbation objective borrowed from PTQ literature; the quadratic approximation may break at 4-bit weights, but the paper provides no validation.
  • domain assumption Activation distributions vary across time steps and time-step-aware calibration captures this variability
    Used in Fig. 3 and in the CQC construction; if false, the quantization calibration set is not representative.
  • ad hoc to paper FP outputs at even teacher time steps align with student outputs at halved time steps
    The time-conditioned uniform sampling in Alg. 1 defines this mapping heuristically; no theoretical or empirical validation beyond the overall results is given.
  • domain assumption Earth-Mover Distance is a computable and effective training loss between student outputs and FP calibration outputs in image space
    EM distance is used in Eq. (8); the paper only ablates loss metrics on one dataset and does not discuss computational cost or high-dimensional EM estimation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PQCAD-DM: Progressive Quantization and Calibration-Assisted Distillation for Extremely Efficient Diffusion Model." pith.science (2026). https://pith.science/paper/GR6YBPH3

@misc{pith2026250616776,
  author       = {Pith},
  title        = {Pith review of: PQCAD-DM: Progressive Quantization and Calibration-Assisted Distillation for Extremely Efficient Diffusion Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GR6YBPH3}},
  note         = {Machine review of arXiv:2506.16776}
}
read the original abstract

Diffusion models excel in image generation but are computational and resource-intensive due to their reliance on iterative Markov chain processes, leading to error accumulation and limiting the effectiveness of naive compression techniques. In this paper, we propose PQCAD-DM, a novel hybrid compression framework combining Progressive Quantization (PQ) and Calibration-Assisted Distillation (CAD) to address these challenges. PQ employs a two-stage quantization with adaptive bit-width transitions guided by a momentum-based mechanism, reducing excessive weight perturbations in low-precision. CAD leverages full-precision calibration datasets during distillation, enabling the student to match full-precision performance even with a quantized teacher. As a result, PQCAD-DM achieves a balance between computational efficiency and generative quality, halving inference time while maintaining competitive performance. Extensive experiments validate PQCAD-DM's superior generative capabilities and efficiency across diverse datasets, outperforming fixed-bit quantization methods.

Figures

Figures reproduced from arXiv: 2506.16776 by the authors.

Figure 1
Figure 1. (Left) Compounding quantization and distillation in DMs introduces cumulative errors at each time step, significantly impacting [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of PQCAD-DM framework. (Left) Dual calibration datasets, (Center) Progressive Quantization (PQ), and (Right) Calibration￾Assisted Distillation (CAD) optimize compression and generative quality [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Activation distribution over time step of diffusion. Figure [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: 256 × 256 image generation results using TFMQ-DM, Q-Diff and PQCAD-DM under 4/8-bit precision. Methods G-Mean Perturb (↓) & ∆Pert(orig−ours)(↑) Bedroom Churches CelebA Q-Diffusion 21.75 22.14 37.40 + PQ (ours) 16.26 (∆311) 17.62 (∆341) 32.95 (∆367) PTQ4DM 25.79 24.93 4…
Figure 5
Figure 5. Figure 5: (Top) Weight perturbation loss over iterations for the [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Changes in LCD during the 500-step distillation process, plotted with respect to distance. Colors represent different distillation stages. Results are based on unconditional image generation with a 256×256 LSUN-Bedrooms LDM at 8/8-bit 100 sampling steps [PITH_FULL_IMA…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 22 canonical work pages

  1. [2]

    Generative adversarial networks,

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio, “Generative adversarial networks,” Communications of the ACM , vol. 63, no. 11, pp. 139–144, 2020

    work page 2020

  2. [3]

    Diffusion models beat gans on image synthesis,

    Prafulla Dhariwal and Alexander Nichol, “Diffusion models beat gans on image synthesis,” Advances in neural information processing systems, vol. 34, pp. 8780–8794, 2021

    work page 2021

  3. [4]

    Auto-encoding variational bayes,

    Diederik P Kingma, “Auto-encoding variational bayes,” arXiv preprint arXiv:1312.6114, 2013

    arXiv 2013

  4. [5]

    Score-based generative modeling through stochastic differential equations,

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole, “Score-based generative modeling through stochastic differential equations,” arXiv preprint arXiv:2011.13456, 2020

    arXiv 2011

  5. [6]

    Repaint: Inpainting using denoising diffu- sion probabilistic models,

    Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool, “Repaint: Inpainting using denoising diffu- sion probabilistic models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 11461–11471

    work page 2022

  6. [7]

    Permutation invariant graph generation via score- based generative modeling,

    Chenhao Niu, Yang Song, Jiaming Song, Shengjia Zhao, Aditya Grover, and Stefano Ermon, “Permutation invariant graph generation via score- based generative modeling,” in International Conference on Artificial Intelligence and Statistics . PMLR, 2020, pp. 4474–4484

    work page 2020

  7. [8]

    Geodiff: A geometric diffusion model for molecular conformation generation,

    Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang, “Geodiff: A geometric diffusion model for molecular conformation generation,” arXiv preprint arXiv:2203.02923 , 2022

    arXiv 2022

  8. [9]

    Post-training quantization on diffusion models,

    Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan, “Post-training quantization on diffusion models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 1972–1981

    work page 2023

Show all 31 references
  1. [10]

    Q-diffusion: Quantizing diffusion models,

    Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer, “Q-diffusion: Quantizing diffusion models,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 17535–17545

    2023

  2. [11]

    Progressive distillation for fast sampling of diffusion models,

    Tim Salimans and Jonathan Ho, “Progressive distillation for fast sampling of diffusion models,” arXiv preprint arXiv:2202.00512, 2022

    2022 arXiv

  3. [12]

    On distillation of guided diffusion models,

    Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans, “On distillation of guided diffusion models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 14297–14306

    2023

  4. [13]

    Denoising diffusion prob- abilistic models,

    Jonathan Ho, Ajay Jain, and Pieter Abbeel, “Denoising diffusion prob- abilistic models,” Advances in neural information processing systems , vol. 33, pp. 6840–6851, 2020

    2020

  5. [14]

    Denoising diffusion implicit models,

    Jiaming Song, Chenlin Meng, and Stefano Ermon, “Denoising diffusion implicit models,” arXiv preprint arXiv:2010.02502 , 2020

    2010 arXiv

  6. [16]

    Improved denoising diffusion probabilistic models,

    Alexander Quinn Nichol and Prafulla Dhariwal, “Improved denoising diffusion probabilistic models,” in International conference on machine learning. PMLR, 2021, pp. 8162–8171

    2021

  7. [17]

    Structural pruning for diffusion models,

    Gongfan Fang, Xinyin Ma, and Xinchao Wang, “Structural pruning for diffusion models,” Advances in Neural Information Processing Systems , vol. 36, 2024

    2024

  8. [18]

    Ld-pruner: Efficient pruning of latent diffusion models using task-agnostic insights,

    Thibault Castells, Hyoung-Kyu Song, Bo-Kyeong Kim, and Shinkook Choi, “Ld-pruner: Efficient pruning of latent diffusion models using task-agnostic insights,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 821–830

    2024

  9. [19]

    Distilling the knowledge in a neural network,

    Geoffrey Hinton, “Distilling the knowledge in a neural network,” arXiv preprint arXiv:1503.02531, 2015

    2015 arXiv

  10. [20]

    Towards effective low-bitwidth convolutional neural networks,

    Bohan Zhuang, Chunhua Shen, Mingkui Tan, Lingqiao Liu, and Ian Reid, “Towards effective low-bitwidth convolutional neural networks,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 7920–7928

    2018

  11. [21]

    Effective training of convolutional neural networks with low-bitwidth weights and activations,

    Bohan Zhuang, Mingkui Tan, Jing Liu, Lingqiao Liu, Ian Reid, and Chunhua Shen, “Effective training of convolutional neural networks with low-bitwidth weights and activations,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 10, pp. 6140–6152, 2021

    2021

  12. [22]

    Adaptive loss-aware quantization for multi-bit networks,

    Zhongnan Qu, Zimu Zhou, Yun Cheng, and Lothar Thiele, “Adaptive loss-aware quantization for multi-bit networks,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2020, pp. 7988–7997

    2020

  13. [23]

    Ctmq: Cyclic training of convolutional neural networks with multiple quantization steps,

    HyunJin Kim, Jungwoo Shin, and Alberto A Del Barrio, “Ctmq: Cyclic training of convolutional neural networks with multiple quantization steps,” arXiv preprint arXiv:2206.12794 , 2022

    2022 arXiv

  14. [24]

    Learning to quantize deep networks by optimizing quantization intervals with task loss,

    Sangil Jung, Changyong Son, Seohyung Lee, Jinwoo Son, Jae-Joon Han, Youngjun Kwak, Sung Ju Hwang, and Changkyu Choi, “Learning to quantize deep networks by optimizing quantization intervals with task loss,” in Proceedings of the IEEE/CVF conference on computer vision and patte...

    2019

  15. [25]

    Bit-shrinking: Limiting instantaneous sharpness for improving post-training quantization,

    Chen Lin, Bo Peng, Zheyang Li, Wenming Tan, Ye Ren, Jun Xiao, and Shiliang Pu, “Bit-shrinking: Limiting instantaneous sharpness for improving post-training quantization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 16196–16205

    2023

  16. [26]

    Vari- ational diffusion models,

    Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho, “Vari- ational diffusion models,” Advances in neural information processing systems, vol. 34, pp. 21696–21707, 2021

    2021

  17. [27]

    Temporal dynamic quantization for diffusion models,

    Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, and Eunhyeok Park, “Temporal dynamic quantization for diffusion models,” Advances in Neural Information Processing Systems , vol. 36, 2024

    2024

  18. [28]

    Tfmq-dm: Temporal feature maintenance quantization for diffusion models,

    Yushi Huang, Ruihao Gong, Jing Liu, Tianlong Chen, and Xianglong Liu, “Tfmq-dm: Temporal feature maintenance quantization for diffusion models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 7362–7371

    2024

  19. [29]

    Up or down? adaptive rounding for post- training quantization,

    Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort, “Up or down? adaptive rounding for post- training quantization,” in International Conference on Machine Learn- ing. PMLR, 2020, pp. 7197–7206

    2020

  20. [30]

    Brecq: Pushing the limit of post-training quantization by block reconstruction,

    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 reconstruction,” arXiv preprint arXiv:2102.05426, 2021

    2021 arXiv

  21. [31]

    High-resolution image synthesis with latent diffu- sion models,

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer, “High-resolution image synthesis with latent diffu- sion models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 10684–10695

    2022

  22. [32]

    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

    2022 arXiv

  23. [33]

    Learned step size quantiza- tion,

    Steven K Esser, Jeffrey L McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S Modha, “Learned step size quantiza- tion,” arXiv preprint arXiv:1902.08153 , 2019. 10 Beomseok Ko (Student Member, IEEE) received the B.S. degree in Information and Communication Eng...

    1902 arXiv

Pith tools

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