Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Efficiency Meets Fidelity: A Novel Quantization Framework for Stable Diffusion

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read The paper proposes a Serial-to-Parallel pipeline for W4A8 quantization that keeps Stable Diffusion outputs close to full-precision images while training faster than prior methods.

desk verdict A useful W4A8 quantization recipe for Stable Diffusion with strong empirical gains, but the headline claim overreaches and the training-inference consistency premise is not actually checked. read the letter →

arxiv 2412.06661 v2 pith:MA3A2XGR submitted 2024-12-09 cs.CV

Shuaiting Li , Juncan Deng , Zeyu Wang , Kedong Xu , Rongtao Deng , Hong Gu , Haibin Shen , Kejie Huang This is my paper · ORCID
classification cs.CV
keywords diffusionmodelquantizationquantization-awaretrainingpost-trainingStabletext-to-imagegenerationedgedeploymentactivationknowledgedistillation
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

This paper argues that Stable Diffusion models can be quantized to 4-bit weights and 8-bit activations without visibly drifting from their floating-point originals, and that this can be done faster than current state-of-the-art quantization methods. The authors identify two failure modes in existing pipelines: serial fine-tuning produces periodic gradient oscillations that destabilize quantized weights, and parallel training uses latents whose numerical range differs from inference-time latents, causing activation misalignment. Their Serial-to-Parallel pipeline avoids both problems by generating latents once with the floating-point model at random timesteps and then training the quantized model in parallel on those stored latents. On SD1.4, SD2.1, SDXL, and SD3, the paper reports FID-to-FP values far below prior methods at W4A8, with shorter training times. If correct, this makes 4-bit deployment of text-to-image models on edge hardware practical while keeping outputs predictable for professional use.

What carries the argument

The central mechanism is the Serial-to-Parallel pipeline. In the serial phase, a floating-point Stable Diffusion model is run on text prompts, and the latent states encountered at randomly selected denoising timesteps are stored; in the parallel phase, the quantized denoising network is trained on these stored latents with the floating-point model as teacher, using an output MSE loss plus feature distillation on selected sensitive layers. This combines the inference-consistent input distribution of serial training with the gradient-averaging stability of parallel training. Around it sit four supporting techniques: separate activation quantization parameters per timestep, precomputation of time-embedding projections so those layers can be removed from the quantized network, distillation targeted at shortcut and projection layers, and iterative freezing of the distilled layers to suppress weight oscillation.

What would settle it

Run the quantized W4A8 model end-to-end over a full 50-step sampling schedule, record its own intermediate latents at each timestep, and compare them to the floating-point model's latents at the same timesteps with the same noise seed and prompt; if the distribution distance grows with step count or the FID-to-FP computed on longer schedules degrades substantially from the reported near-FP values, the Serial-to-Parallel premise would be contradicted.

Watch

Extended reading notes

Core claim

The central claim is that quantized Stable Diffusion models can reproduce floating-point generation with high fidelity under W4A8 quantization when training and inference are made consistent. The paper's key evidence is FID-to-FP, the distribution distance between images from the quantized model and images from the full-precision model. For SD1.4 on COCO prompts, W4A8 FID-to-FP drops from 14.2 (PCR) to 9.46, with SSIM rising from 0.47 to 0.63; for SDXL it drops from 18.27 to 13.35; for SD3 at 1024x1024 the method reaches 9.64, well below NF4 baselines that use 16-bit activations. The authors conclude that a combination of the Serial-to-Parallel pipeline, per-timestep activation quantizers, precalculated time embeddings, inter-layer distillation on sensitive layers, and selective freezing yields images that match the floating-point model in layout, color, and style, not merely in quality.

Load-bearing premise

The approach assumes that the latents stored from a floating-point model's denoising run are representative of the latents the quantized model will see when it denoises on its own; if accumulated quantization error pushes the quantized model's internal latents off that distribution, the training-inference consistency the pipeline is built on weakens.

Editorial extensions

If this is right

  • At W4A8, quantized SD1.4, SD2.1, SDXL, and SD3 can produce images whose distribution and pixel structure closely track the FP16 model, so professional pipelines can swap in quantized models without re-tuning prompts.
  • Training time is much lower than the PCR baseline (about 4.5h versus 13h for SD1.4; 7.5h versus 25h for SDXL) because latent dataset generation is done once and reused.
  • Because time-embedding and time-projection layers no longer need to be quantized, one source of image-quality degradation is removed at negligible memory cost.
  • The pipeline is data-free, relying only on prompts, and extends to architectures like MM-DiT-based SD3, where the paper shows W4A8 beating NF4 W4A16.
  • The per-timestep activation quantizers can be optimized simultaneously in one batch, making training efficient while keeping activation ranges accurate across the denoising schedule.

Reading between the lines

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

  • Because the paper's fidelity numbers compare against floating-point-generated latents, a natural extension is to measure FID-to-FP over the quantized model's own multistep denoising trajectories; fixed per-timestep activation scales may need re-calibration if quantization error accumulates across steps.
  • The two failure modes the paper identifies, gradient oscillation in serial training and latent-range mismatch in parallel training, are generic to iterative generative models, so the Serial-to-Parallel recipe could transfer to video diffusion or other transformer-based denoisers.
  • Precomputing time features means the quantized model no longer needs to compute time embeddings at inference, which adds a small extra latency and memory saving beyond the numbers reported.
  • A cheap diagnostic test for practitioners is to compare the quantized model's self-generated latents with the floating-point model's latents under the same prompt and noise seed; the divergence rate would forecast how much style drift appears on longer runs.
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 / 6 minor

Summary. The paper proposes a quantization-aware fine-tuning framework for Stable Diffusion models aimed at preserving fidelity to the floating-point generator. The core idea is a Serial-to-Parallel pipeline: latents are first cached by running the FP model for random numbers of denoising steps, and these precomputed latents are then used for parallel distillation-style training of the quantized UNet. On top of this, the authors add per-timestep activation quantizers, precomputed time embeddings, inter-layer distillation on selected sensitive layers, and selective freezing of oscillating layers. Experiments on SD1.4, SD2.1, SDXL, and SD3 under W4A8/W8A8 quantization report FID-to-FP, SSIM, Image Reward, and CLIP scores, claiming large FID-to-FP improvements over PCR and shorter training times.

Significance. If the reported results hold, this is a practically valuable contribution: it suggests that 4-bit quantized Stable Diffusion variants can closely match the outputs of their FP counterparts, which matters for edge deployment and for professional workflows that require predictable, consistent generation. The paper has clear strengths: the motivating observations about serial vs. parallel training are concrete; the ablation in Table 5 shows each component adds a measurable improvement; the efficiency comparison in Table 4 is favorable; and the evaluation spans four model families, including an MMDiT-based model. The main weakness is that the central consistency claim is validated only under one sampling configuration, and several experiments lack the reproducibility details (code, seeds, error bars) needed to fully trust the magnitude of the reported gains.

major comments (4)
  1. [4.1, Fig. 5, Eq. (6)] The Serial-to-Parallel pipeline trains the quantized model on latents cached from the floating-point model, but at deployment the input to the quantized UNet at each step is produced by the quantized model's own previous step. Observation 2 and Eq. (6) only contrast FP training latents with FP inference latents; they never verify that the quantized model's self-generated latents match the FP latent distribution. Since the activation quantizer parameters are calibrated on FP latents, accumulated quantization error could shift the deployment distribution and invalidate the claimed training-inference consistency. The reported FID-to-FP under the default scheduler is indirect evidence, but it covers a single configuration. Please add a direct comparison of per-timestep latent statistics (or a distributional distance) between FP and quantized self-generated latents, and report FID-to-FP for at least two different schedulers, step counts, and CFG scales.
  2. [Table 2, SD1.4 W8A8 row] For SD1.4 COCO at W8A8, the proposed method reports FID-to-FP 8.57, which is worse than PCR's 8.35 on the key similarity metric, even though the text states that 'our method dominates other metrics.' The abstract and introduction claim superior performance over state-of-the-art approaches without restricting the claim to W4A8. Please either explicitly scope the headline claim to the W4A8 setting, or provide an explanation or error bars showing that the W8A8 difference is within noise.
  3. [5.2, Table 2, SD3 rows] The SD3 comparison uses NF4 (W4A16) as the only baseline. NF4 is a generic weight-only quantization method from LLM fine-tuning, not a diffusion-specific PTQ method, so it does not substantiate the claim of superiority over state-of-the-art quantization approaches for SD3. Since no diffusion-specific PTQ baseline is adapted to SD3, the SD3 results currently support scalability of the pipeline but not the 'superior performance over SOTA' conclusion. Please either adapt a diffusion-specific method (e.g., Q-diffusion or PCR) to SD3, or explicitly present the SD3 comparison as a baseline-free feasibility study.
  4. [5.1 (Implementation details)] Reproducibility is hampered by missing training details: the paper does not specify the optimizer, learning rate, batch size, number of training iterations, exact subset sizes aside from prompt counts, or random seeds for any of the main results. Given that efficiency and shorter training time are central claims, these details are needed to assess both the validity and the variability of the reported numbers. Please add a complete hyperparameter table and, if possible, run each main configuration with at least three seeds and report standard deviations for FID-to-FP.
minor comments (6)
  1. [Abstract] There are capitalization and grammar issues: 'While Recent studies' should be 'While recent studies,' and the first sentence uses 'Stable Diffusion models (SDM)' with a plural verb but a singular acronym; please correct throughout.
  2. [Eq. (6)] Equation (6) conflates a model output with a random variable: the left side is the predicted mean, while the right side is a conditional distribution. Please rewrite the equation to compare distributions of latents rather than mixing a deterministic function with a sampling expression.
  3. [Table 2] The table header 'W8A8.4' and 'W4A8.4' is confusing; the text explains that PCR relaxes 20% of timesteps to 10-bit, but this should be stated directly in the table caption. Also, the SD3 NF4 rows use W4A16, which should be clarified in the caption as weight-only quantization.
  4. [Table 2] Several cells use inconsistent formatting: '26,33' should be '26.33', 'Clip' should be 'CLIP', and the SD-XL resolution column says '768' but the text uses 'SDXL'; please standardize.
  5. [5.3, Table 3] The dataset generation comparison reports FID-to-FP for 50 steps/prompt and 1 step/prompt, but it does not report how many timesteps were sampled per latent in the 1-step case beyond the name; please clarify whether '1 step' means one random timestep per prompt or a full single denoising step from the initial noise.
  6. [Figures 7 and 8] The qualitative figures would benefit from a layout that aligns each method column with the same prompt row and from captions describing what visual differences the reader should look for; currently the figures are referenced without detailed discussion of failure modes.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the training objective is standard distillation and the headline fidelity metrics are evaluated externally, not by construction.

full rationale

The paper's derivation chain is empirical rather than definitional. The load-bearing objective (Eq. 9) minimizes the MSE between quantized and floating-point DDB outputs on floating-point-generated latents; this is standard output distillation. The reported similarity metric, FID-to-FP, is computed on full generated images in Inception feature space, so it is not identical to the training loss by construction. The framework's other evidence (IR, CLIP score, SSIM, and comparisons to PCR/NF4 baselines) uses external evaluators that are independent of the training loss. The Serial-to-Parallel pipeline does rest on an empirical assumption that FP-generated training latents are representative of inference-time latents, and the paper itself acknowledges the FP-training/inference mismatch (Observation 2, Eq. 6); however, this is a distributional and robustness risk, not circular reasoning, because the claimed results are not derived from that assumption. No author self-citations are used to justify the central result; prior work is cited for baselines and component techniques. No equation reduces to its own input, and no fitted parameter is renamed as a prediction, so no circular step is exhibited.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The framework is primarily empirical; the main added degrees of freedom are the per-timestep activation quantizers and the hand-chosen dataset, freezing, and layer-selection hyperparameters. No new physical or conceptual entities are introduced.

free parameters (4)
  • per-timestep activation scale/zero-point sets = T sets per quantized layer
    Eq. 7 introduces separate s_l^t and z_l^t for each timestep; these are learned from calibration data and are additional degrees of freedom beyond a single global activation quantizer.
  • latent dataset size = 20,000 prompts (SD1.4), 10,000 prompts (SDXL)
    Chosen after comparing 4,000 prompts x 50 latents vs 20,000 prompts x 1 latent (Tab. 3) and observing overfitting in Fig. 9. This selection directly affects FID-to-FP results.
  • freezing frequency = every 500 iterations
    Set relative to Osqat's per-iteration freezing; no sensitivity analysis is reported (Section 4.2, Selective Freezing).
  • sensitive layer set = shortcut + FFN projection layers (UNet); FFN projection layers (MMDiT)
    Hand-selected based on prior work [29,49,53], not on an analysis of the models used here.
assumptions (6)
  • standard math Diffusion denoising is approximated by the DDPM/DDIM formulations in Section 3.1, including the backward sampling equation (Eq. 3).
    The framework inherits the standard stochastic/ODE sampling assumptions of Stable Diffusion; no alternative generative model is considered.
  • domain assumption Minimizing per-step output MSE between quantized and floating-point DDB (Eq. 9) is a sufficient surrogate for perceptual and distributional consistency of final images.
    This is the core training objective; the evaluation on FID-to-FP, SSIM, IR, and CLIP provides empirical support but the equivalence is not proven.
  • ad hoc to paper Latents sampled once from the FP model at random timesteps remain representative of the inputs the quantized model will see during inference.
    Section 4.1 Observation 2 motivates using FP latents but does not analyze drift caused by quantized self-generation; this is the paper's weakest premise.
  • domain assumption Periodic gradient oscillations in serial training cause unstable quantized weights and must be avoided.
    Figures 2 and 3 and citations [37,40,57] support this, but the causal chain is inferred from correlation.
  • domain assumption Activation statistics differ enough across timesteps that per-timestep quantizers are necessary.
    Based on prior work [24,45,46,53,60]; no distribution statistics for the tested models are shown.
  • standard math Time projection outputs e_p are deterministic functions of timestep t only (Eq. 8), so precomputing them is exact.
    True by definition of the embedding and projection layers; not a source of error.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficiency Meets Fidelity: A Novel Quantization Framework for Stable Diffusion." pith.science (2026). https://pith.science/paper/MA3A2XGR

@misc{pith2026241206661,
  author       = {Pith},
  title        = {Pith review of: Efficiency Meets Fidelity: A Novel Quantization Framework for Stable Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MA3A2XGR}},
  note         = {Machine review of arXiv:2412.06661}
}
read the original abstract

Text-to-image generation via Stable Diffusion models (SDM) have demonstrated remarkable capabilities. However, their computational intensity, particularly in the iterative denoising process, hinders real-time deployment in latency-sensitive applications. While Recent studies have explored post-training quantization (PTQ) and quantization-aware training (QAT) methods to compress Diffusion models, existing methods often overlook the consistency between results generated by quantized models and those from floating-point models. This consistency is paramount for professional applications where both efficiency and output reliability are essential. To ensure that quantized SDM generates high-quality and consistent images, we propose an efficient quantization framework for SDM. Our framework introduces a Serial-to-Parallel pipeline that simultaneously maintains training-inference consistency and ensures optimization stability. Building upon this foundation, we further develop several techniques including multi-timestep activation quantization, time information precalculation, inter-layer distillation, and selective freezing, to achieve high-fidelity generation in comparison to floating-point models while maintaining quantization efficiency. Through comprehensive evaluation across multiple Stable Diffusion variants (v1-4, v2-1, XL 1.0, and v3), our method demonstrates superior performance over state-of-the-art approaches with shorter training times. Under W4A8 quantization settings, we achieve significant improvements in both distribution similarity and visual fidelity, while preserving a high image quality.

Figures

Figures reproduced from arXiv: 2412.06661 by the authors.

Figure 2
Figure 2. Box plot illustrating the gradient variations in the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Comparison of loss (left) and weight oscillation [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Difference in noisy input range at each timestep [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Overview of our quantization framework. (A) Serial dataset generation: During the inference of floating-point models, [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Weight oscillations in selected sensitive layers. The [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Stable Diffusion v1.4 512×512 image generation using COCO prompts and Stable-Diffusion-Prompts [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Stable Diffusion v3 1024×1024 image generation using COCO prompts and Stable-Diffusion-Prompts. weight quantization and static per-tensor activation quantization. Except for special declaration, we adopt the default noise sched￾uler and Classifier-Free Guidance (CFG) s…
Figure 9
Figure 9. Figure 9: Comparison of Loss and FID-to-FP Curves under [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Smaller Models, Unexpected Costs: Trade-offs in LLM Quantization for Automated Program Repair

    cs.SE 2026-06 unverdicted novelty 6.0 of 10

    Empirical evaluation of 13 quantization configurations on 6 LLMs for APR shows reduced memory (up to 85%) but increased inference time/energy, different repaired problem sets with little overlap, and 48% of configs st...

Reference graph

Works this paper leans on

72 extracted references · 18 canonical work pages · cited by 1 Pith paper

  1. [1]

    Fan Bao, Chongxuan Li, Jiacheng Sun, Jun Zhu, and Bo Zhang. 2022. Estimating the Optimal Covariance with Imperfect Mean in Diffusion Probabilistic Models. arXiv:2206.07309 [cs.LG] https://arxiv.org/abs/2206.07309

    arXiv 2022

  2. [2]

    Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. 2022. Analytic-DPM: an Analytic Estimate of the Optimal Reverse Variance in Diffusion Probabilistic Models. arXiv:2201.06503 [cs.LG] https://arxiv.org/abs/2201.06503

    arXiv 2022

  3. [3]

    Chaofeng Chen and Jiadi Mo. 2022. IQA-PyTorch: PyTorch Toolbox for Image Quality Assessment. [Online]. Available: https://github.com/chaofengc/IQA- PyTorch

    work page 2022

  4. [4]

    Hong Chen, Yipeng Zhang, Simin Wu, Xin Wang, Xuguang Duan, Yuwei Zhou, and Wenwu Zhu. 2024. DisenBooth: Identity-Preserving Disentangled Tuning for Subject-Driven Text-to-Image Generation. arXiv:2305.03374 [cs.CV] https: //arxiv.org/abs/2305.03374

    arXiv 2024

  5. [5]

    Hyungjin Chung, Byeongsu Sim, and Jong Chul Ye. 2022. Come-closer-diffuse- faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 12413–12422. Efficiency Meets Fidelity: A Novel Quantization Framework for Stable Diffusion ...

    work page 2022

  6. [6]

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. Qlora: Efficient finetuning of quantized llms. Advances in neural information processing systems 36 (2023), 10088–10115

    2023

  7. [7]

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

    2021

  8. [8]

    Giulio Franzese, Simone Rossi, Lixuan Yang, Alessandro Finamore, Dario Rossi, Maurizio Filippone, and Pietro Michiardi. 2023. How much is enough? a study on diffusion times in score-based generative models. Entropy 25, 4 (2023), 633

    work page 2023

Show all 72 references
  1. [9]

    Ruihao Gong, Xianglong Liu, Shenghu Jiang, Tianxiang Li, Peng Hu, Jiazhen Lin, Fengwei Yu, and Junjie Yan. 2019. Differentiable soft quantization: Bridg- ing full-precision and low-bit neural networks. In Proceedings of the IEEE/CVF international conference on computer vision ...

    2019

  2. [10]

    Yefei He, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. 2023. Efficientdm: Efficient quantization-aware fine-tuning of low-bit diffusion models. arXiv preprint arXiv:2310.03270 (2023)

    2023 arXiv

  3. [11]

    Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. 2024. Ptqd: Accurate post-training quantization for diffusion models. Advances in Neural Information Processing Systems 36 (2024)

    2024

  4. [12]

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. 2021. Clipscore: A reference-free evaluation metric for image captioning.arXiv preprint arXiv:2104.08718 (2021)

    2021 arXiv

  5. [13]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models. Advances in neural information processing systems 33 (2020), 6840–6851

    2020

  6. [14]

    Yushi Huang, Ruihao Gong, Jing Liu, Tianlong Chen, and Xianglong Liu. 2024. TFMQ-DM: Temporal Feature Maintenance Quantization for Diffusion Models. arXiv:2311.16503 [cs.CV] https://arxiv.org/abs/2311.16503

    2024 arXiv

  7. [15]

    Itay Hubara, Yury Nahshan, Yair Hanani, Ron Banner, and Daniel Soudry. 2020. Improving Post Training Neural Quantization: Layer-wise Calibration and Integer Programming. arXiv:2006.10518 [cs.LG] https://arxiv.org/abs/2006.10518

    2020 arXiv

  8. [16]

    Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. 2018. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In Proceedings of the IEEE conference on computer vis...

    2018

  9. [17]

    Alexia Jolicoeur-Martineau, Ke Li, Rémi Piché-Taillefer, Tal Kachman, and Ioannis Mitliagkas. 2021. Gotta Go Fast When Generating Data with Score-Based Models. arXiv:2105.14080 [cs.LG] https://arxiv.org/abs/2105.14080

    2021 arXiv

  10. [18]

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. 2022. Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems 35 (2022), 26565–26577

    2022

  11. [19]

    Beomsu Kim and Jong Chul Ye. 2022. Denoising MCMC for Accelerating Diffusion-Based Generative Models. arXiv:2209.14593 [cs.LG] https://arxiv. org/abs/2209.14593

    2022 arXiv

  12. [20]

    Bo-Kyeong Kim, Hyoung-Kyu Song, Thibault Castells, and Shinkook Choi. 2023. Bk-sdm: Architecturally compressed stable diffusion for efficient text-to-image generation. In Workshop on Efficient Systems for Foundation Models@ ICML2023

    2023

  13. [21]

    Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. 2021. Variational diffusion models. Advances in neural information processing systems 34 (2021), 21696–21707

    2021

  14. [22]

    Zhifeng Kong and Wei Ping. 2021. On Fast Sampling of Diffusion Probabilistic Models. arXiv:2106.00132 [cs.LG] https://arxiv.org/abs/2106.00132

    2021 arXiv

  15. [23]

    Max W. Y. Lam, Jun Wang, Dan Su, and Dong Yu. 2022. BDDM: Bilat- eral Denoising Diffusion Models for Fast and High-Quality Speech Synthesis. arXiv:2203.13508 [eess.AS] https://arxiv.org/abs/2203.13508

    2022 arXiv

  16. [24]

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

    2023

  17. [25]

    Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. 2021. BRECQ: Pushing the Limit of Post-Training Quantization by Block Reconstruction. arXiv:2102.05426 [cs.LG] https://arxiv. org/abs/2102.05426

    2021 arXiv

  18. [26]

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proce...

    2014

  19. [27]

    Yang Lin, Tianyu Zhang, Peiqin Sun, Zheng Li, and Shuchang Zhou. 2023. FQ-ViT: Post-Training Quantization for Fully Quantized Vision Transformer. arXiv:2111.13824 [cs.CV] https://arxiv.org/abs/2111.13824

    2023 arXiv

  20. [28]

    Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. 2022. Pseudo Numerical Methods for Diffusion Models on Manifolds. arXiv:2202.09778 [cs.CV] https://arxiv.org/ abs/2202.09778

    2022 arXiv

  21. [29]

    Xinyan Liu, Huihong Shi, Yang Xu, and Zhongfeng Wang. 2024. TaQ-DiT: Time- aware Quantization for Diffusion Transformers. arXiv preprint arXiv:2411.14172 (2024)

    2024

  22. [30]

    Christos Louizos, Matthias Reisser, Tijmen Blankevoort, Efstratios Gavves, and Max Welling. 2018. Relaxed Quantization for Discretized Neural Networks. arXiv:1810.01875 [cs.LG] https://arxiv.org/abs/1810.01875

    2018 arXiv

  23. [31]

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu

  24. [32]

    Eric Luhman and Troy Luhman. 2021. Knowledge Distillation in Iterative Generative Models for Improved Sampling Speed. arXiv:2101.02388 [cs.LG] https://arxiv.org/abs/2101.02388

    2021 arXiv

  25. [33]

    Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. 2023. Latent Con- sistency Models: Synthesizing High-Resolution Images with Few-Step Inference. arXiv:2310.04378 [cs.CV] https://arxiv.org/abs/2310.04378

    2023 arXiv

  26. [34]

    Zhaoyang Lyu, Xudong XU, Ceyuan Yang, Dahua Lin, and Bo Dai. 2022. Accelerating Diffusion Models via Early Stop of the Diffusion Process. arXiv:2205.12524 [cs.CV] https://arxiv.org/abs/2205.12524

    2022 arXiv

  27. [35]

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

    2020

  28. [36]

    Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yelysei Bondarenko, Mart van Baalen, and Tijmen Blankevoort. 2021. A White Paper on Neural Network Quantization. arXiv:2106.08295 [cs.LG] https://arxiv.org/abs/2106.08295

    2021 arXiv

  29. [37]

    Markus Nagel, Marios Fournarakis, Yelysei Bondarenko, and Tijmen Blankevoort

  30. [38]

    Alexander Quinn Nichol and Prafulla Dhariwal. 2021. Improved denoising diffu- sion probabilistic models. In International conference on machine learning . PMLR, 8162–8171

    2021

  31. [39]

    In International Conference on Machine Learning

    Overcoming oscillations in quantization-aware training. In International Conference on Machine Learning . PMLR, 16318–16330

  32. [40]

    Sashank J Reddi, Satyen Kale, and Sanjiv Kumar. 2019. On the convergence of adam and beyond. arXiv preprint arXiv:1904.09237 (2019)

    2019 arXiv

  33. [41]

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. 2023. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. arXiv:2307.01952 [cs.CV] https://arxiv.org/abs/2307.01952

    2023 arXiv

  34. [42]

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-net: Convolu- tional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceeding...

    2015

  35. [43]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 10684–10695

    2022

  36. [44]

    Tim Salimans and Jonathan Ho. 2022. Progressive Distillation for Fast Sampling of Diffusion Models. arXiv:2202.00512 [cs.LG] https://arxiv.org/abs/2202.00512

    2022 arXiv

  37. [45]

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. 2023. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 22500–22510

    2023

  38. [46]

    Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, and Eunhyeok Park

  39. [47]

    Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. 2023. Post- training quantization on diffusion models. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition . 1972–1981

    2023

  40. [48]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. 2021. Score-Based Generative Modeling through Stochas- tic Differential Equations. arXiv:2011.13456 [cs.LG] https://arxiv.org/abs/2011. 13456

    2021 arXiv

  41. [49]

    Yang Sui, Yanyu Li, Anil Kag, Yerlan Idelbayev, Junli Cao, Ju Hu, Dhritiman Sagar, Bo Yuan, Sergey Tulyakov, and Jian Ren. 2024. Bitsfusion: 1.99 bits weight quantization of diffusion model. arXiv preprint arXiv:2406.04333 (2024)

    2024 arXiv

  42. [50]

    Jiaming Song, Chenlin Meng, and Stefano Ermon. 2022. Denoising Diffusion Implicit Models. arXiv:2010.02502 [cs.LG] https://arxiv.org/abs/2010.02502

    2022 arXiv

  43. [51]

    Jiayan Teng, Wendi Zheng, Ming Ding, Wenyi Hong, Jianqiao Wangni, Zhuoyi Yang, and Jie Tang. 2023. Relay Diffusion: Unifying diffusion process across resolutions for image synthesis. arXiv:2309.03350 [cs.CV] https://arxiv.org/abs/ 2309.03350

    2023 arXiv

  44. [52]

    Changyuan Wang, Ziwei Wang, Xiuwei Xu, Yansong Tang, Jie Zhou, and Jiwen Lu. 2024. Towards Accurate Post-training Quantization for Diffusion Models. arXiv:2305.18723 [cs.CV] https://arxiv.org/abs/2305.18723

    2024 arXiv

  45. [53]

    Siao Tang, Xin Wang, Hong Chen, Chaoyu Guan, Zewen Wu, Yansong Tang, and Wenwu Zhu. 2024. Post-training Quantization for Text-to-Image Diffusion Models with Progressive Calibration and Activation Relaxing. arXiv:2311.06322 [cs.CV] https://arxiv.org/abs/2311.06322

    2024 arXiv

  46. [54]

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. 2004. Image quality assessment: From error visibility to structural similarity.IEEE Transactions on Image Processing 13, 4 (2004), 600–612

    2004

  47. [55]

    Daniel Watson, Jonathan Ho, Mohammad Norouzi, and William Chan

  48. [56]

    Haoxuan Wang, Yuzhang Shang, Zhihang Yuan, Junyi Wu, and Yan Yan. 2024. QuEST: Low-bit Diffusion Model Quantization via Efficient Selective Finetuning. arXiv:2402.03666 [cs.CV] https://arxiv.org/abs/2402.03666 Conference’17, July 2017, Washington, DC, USA Shuaiting Li et al

    2024 arXiv

  49. [57]

    Ashia C Wilson, Rebecca Roelofs, Mitchell Stern, Nati Srebro, and Benjamin Recht

  50. [58]

    Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. 2024. Seesr: Towards semantics-aware real-world image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recogni- tion. 25456–25467

    2024

  51. [59]

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. 2023. Imagereward: Learning and evaluating human prefer- ences for text-to-image generation. Advances in Neural Information Processing Systems 36 (2023), 15903–15935

    2023

  52. [60]

    Xiuying Wei, Ruihao Gong, Yuhang Li, Xianglong Liu, and Fengwei Yu. 2023. QDrop: Randomly Dropping Quantization for Extremely Low-bit Post-Training Quantization. arXiv:2203.05740 [cs.CV] https://arxiv.org/abs/2203.05740

    2023 arXiv

  53. [61]

    Luoming Zhang, Yefei He, Zhenyu Lou, Xin Ye, Yuxing Wang, and Hong Zhou

  54. [62]

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. 2023. Adding conditional con- trol to text-to-image diffusion models. InProceedings of the IEEE/CVF International Conference on Computer Vision . 3836–3847

    2023

  55. [63]

    Qinsheng Zhang and Yongxin Chen. 2023. Fast Sampling of Diffusion Models with Exponential Integrator. arXiv:2204.13902 [cs.LG] https://arxiv.org/abs/2204. 13902

    2023 arXiv

  56. [64]

    Qinsheng Zhang, Molei Tao, and Yongxin Chen. 2023. gDDIM: Generalized denoising diffusion implicit models. arXiv:2206.05564 [cs.LG] https://arxiv.org/ abs/2206.05564

    2023 arXiv

  57. [65]

    Yuewei Yang, Xiaoliang Dai, Jialiang Wang, Peizhao Zhang, and Hongbo Zhang. 2023. Efficient Quantization Strategies for Latent Diffusion Models. arXiv:2312.05431 [cs.CV] https://arxiv.org/abs/2312.05431

    2023 arXiv

  58. [66]

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

    2018

  59. [71]

    Huangjie Zheng, Pengcheng He, Weizhu Chen, and Mingyuan Zhou. 2023. Truncated Diffusion Probabilistic Models and Diffusion-based Adversarial Auto- Encoders. arXiv:2202.09671 [stat.ML] https://arxiv.org/abs/2202.09671

    2023 arXiv

  60. [2017]

    Advances in neural information processing systems 30 (2017)

    The marginal value of adaptive gradient methods in machine learning. Advances in neural information processing systems 30 (2017)

    2017

  61. [2021]

    arXiv:2106.03802 [cs.LG] https://arxiv.org/abs/2106.03802

    Learning to Efficiently Sample from Diffusion Probabilistic Models. arXiv:2106.03802 [cs.LG] https://arxiv.org/abs/2106.03802

    arXiv

  62. [2022]

    Advances in Neural Information Processing Systems 35 (2022), 5775–5787

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems 35 (2022), 5775–5787

    2022

  63. [2023]

    Applied Intelligence 53, 6 (2023), 6266–6275

    Root quantization: a self-adaptive supplement STE. Applied Intelligence 53, 6 (2023), 6266–6275

    2023

  64. [2024]

    Advances in Neural Information Processing Systems 36 (2024)

    Temporal dynamic quantization for diffusion models. Advances in Neural Information Processing Systems 36 (2024)

    2024

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.