Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Q-Sched: Pushing the Boundaries of Few-Step Diffusion Models with Quantization-Aware Scheduling

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Q-Sched claims that a quantized few-step diffusion model can match or beat full-precision image quality by tuning only two scalar coefficients in the sampling schedule, achieving a 4× model-size reduction without retraining weights.

desk verdict A simple, reference-free scheduler tweak that gives real gains for 4-bit few-step diffusion, but the paper overstates the guarantee and skimps on reproducibility details. read the letter →

arxiv 2509.01624 v1 pith:QEYLI6V6 submitted 2025-09-01 cs.CV cs.AI

Natalia Frumkin , Diana Marculescu This is my paper
classification cs.CVcs.AI
keywords post-trainingquantizationfew-stepdiffusionquantization-awareschedulingconsistencymodelstrajectorydistillationreference-freelosstext-to-imagegenerationmodelcompression
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

The paper tries to establish that quantization noise in few-step diffusion models can be corrected at the scheduler level rather than the weight level. It proposes Q-Sched, which learns two scalar preconditioning coefficients—one scaling the previous latent, one scaling the quantized noise prediction—and optimizes them with a reference-free loss called JAQ that pairs text-image alignment with image quality. If correct, quantized few-step models reach full-precision FID while using a quarter of the memory, and quantization complements rather than fights few-step distillation. The paper reports FID gains of 15.5% over a full-precision 4-step Latent Consistency Model and 16.6% over an 8-step Phased Consistency Model, plus a human-preference study with more than 80,000 annotations on FLUX.1[schnell] and SDXL-Turbo.

What carries the argument

The central object is the modified TCD Strategic Stochastic Sampling update with two learnable scalar preconditioning coefficients, cx and c_epsilon (Eq. 11). The coefficients reweight the two inputs the sampler receives from the previous step—the partially denoised latent xt and the quantized noise prediction E^Q_theta—effectively steering the sampling trajectory away from artifacts induced by quantization and by the few-step distillation itself. The selection mechanism is the JAQ loss, a linear combination of a text-image compatibility metric and a no-reference image quality metric, which serves as a calibration objective that needs no full-precision model or reference images; in practice

What would settle it

Run the same grid search on a 4-step LCM at W3A8 or on a 1-step model and check whether any (cx,c_epsilon) yields FID at or below the full-precision few-step model; if the best pair cannot beat naive quantization, the two-scalar correction is not generally sufficient. Alternatively, compute the optimal coefficients separately on COCO-30k prompts and MJHQ prompts for the same backbone and compare: a large shift in the optimum indicates the correction is prompt-distribution-specific rather than a property of the quantized sampler.

Watch

Extended reading notes

Core claim

The central claim is that compressing the denoising backbone of a few-step diffusion model to W4A8 and then re-tuning the sampler's two scalar coefficients restores, and in some configurations exceeds, full-precision generation quality. The two coefficients enter the TCD strategic stochastic sampling update: cx scales the previous latent xt and c_epsilon scales the quantized noise estimate E^Q_theta, so the corrected step is xs = (alpha_s/alpha_s')[(alpha_s'/alpha_t) cx xt - sigma_t c_epsilon E^Q_theta + sigma_s' c_epsilon E^Q_theta] + sqrt(1 - alpha_s^2/alpha_s'^2) z. The coefficients are found by grid search minimizing JAQ(x) = TC(x) + k·IQ(x), a reference-free sum of a text-image compatib

Load-bearing premise

One pair of scalar coefficients, applied to the previous latent and the quantized noise prediction, can correct the quantization-induced trajectory shift well enough for every model, bit-width, step count, and prompt distribution the method is used on.

Editorial extensions

If this is right

  • W4A8 quantized few-step models can replace FP16 ones in deployment, cutting backbone memory by 4× while keeping or improving FID, so text-to-image models become practical on consumer GPUs.
  • Post-training quantization of diffusion models no longer needs a large calibration set of full-precision outputs: five to twenty prompts and a reference-free loss suffice.
  • Quantization and few-step distillation can be combined as complementary compression strategies rather than treated as competing ones.
  • The scheduler-level correction is backbone-agnostic, applying to both U-Net and DiT architectures via the same two-scalar recipe.
  • The JAQ loss offers a reusable reference-free objective for tuning generative model hyperparameters beyond scheduling, wherever ground truth images are unavailable.

Reading between the lines

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

  • Because the correction is only two global scalars, there is likely headroom to generalize it to per-timestep or per-block coefficients for harder regimes (W3A8, W4A4, 1-step models), where a single pair may not absorb the accumulated error.
  • The success of trajectory-level correction suggests that quantization error in few-step diffusion behaves less like additive noise and more like a smooth re-parametrization of the ODE path; if so, the same idea could transfer to other latent-space perturbations such as LoRA merges or activation sparsity.
  • The reported W8A8 case on SDXL-Turbo where Q-Sched's FID trailed the baselines hints that the two-scalar correction helps most when quantization actually shifts the trajectory; at higher precision the added degrees of freedom may overfit the small calibration set.
  • A direct test would be to measure how much the optimal (cx, c_epsilon) moves when the prompt distribution changes; if the optimum is prompt-specific, the method needs a per-domain calibration step in practice.
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 / 5 minor

Summary. The paper proposes Q-Sched, a post-training quantization method for few-step text-to-image diffusion models. Instead of modifying the network weights, Q-Sched adjusts the sampling trajectory by introducing two scalar pre-conditioning coefficients, c_x and c_epsilon, into the TCD-style scheduler (Eq. 11). These coefficients are selected by grid search to minimize a reference-free JAQ loss (Eq. 12), which combines a text-image compatibility term and an image quality term. The method is evaluated on LCM and PCM models with a Stable Diffusion v1.5 backbone (Table 1), on a 2-step PCM with SDXL (Table 2a), and on SDXL-Turbo and FLUX.1[schnell] at multiple bit-widths (Table 2b), with FID, CLIPScore, and large-scale human preference annotations. The authors claim that Q-Sched achieves full-precision accuracy with a 4x reduction in model size and that quantization and few-step distillation are complementary.

Significance. If the empirical claims hold, Q-Sched is an attractive and modular idea: it requires no full-precision calibration, uses only a handful of prompts, and can be applied on top of existing quantization methods. The paper includes a broad set of experiments, including LCM, PCM, SDXL-Turbo, and FLUX.1, and a large-scale user study with more than 80,000 annotations, which is a substantial strength. The central FID-based claims are generally supported by the tables in the lower-bit (W4A8) regimes. However, the paper overgeneralizes in places: the abstract and several statements are contradicted by the W8A8 result in Table 2b, and the theoretical appendix claiming a 'strict guarantee' is mathematically unsound. The absence of error bars, learned coefficient values, grid-search ranges, and the calibration prompt list further limits the reproducibility and strength of the empirical conclusions.

major comments (4)
  1. [Abstract / Table 2b] The abstract states unconditionally that Q-Sched 'achieves full-precision accuracy with a 4x reduction in model size.' However, Table 2b reports SDXL-Turbo W8A8 with Q-Sched FID 26.34, which is worse than Naive W8A8 (25.49), MixDQ W8A8 (25.16), and FP16 (25.48). In this setting, the proposed scheduler is worse than doing nothing. The main text later says 'advantages are most pronounced in lower-bit regimes,' but the abstract and the sentence 'W4A8 and W8A8 ... yet our method improves over baseline' do not carry this scope. Please restrict the claims to the tested W4A8 regime, or provide a clear explanation and a mechanism that avoids this failure.
  2. [Section 3 / Appendix H] The central assertion that two scalar coefficients are sufficient is not established. Appendix H's 'strict guarantee' is invalid: Eq. (33) writes E[||Δx0||] as a sum of terms E[||ΔEθ(s)||], but the norm of a sum is not the sum of norms without independence or orthogonality assumptions that are not stated. Moreover, Q-Sched uses global scalars c_x and c_epsilon, while Appendix H analyzes per-step coefficients k_t and m_t; the existence argument for per-step coefficients does not apply to the actual parametrization. The W8A8 result in Table 2b is a concrete counterexample. I recommend either proving a scoped guarantee for the actual two-scalar form, removing Appendix H entirely, or clearly labeling the two-scalar sufficiency as an empirical observation limited to the tested settings.
  3. [Section 4 / Eq. (12)] For LCM and PCM, the JAQ loss uses TC(x) = CLIPScore(x), and the same metric is then reported as an evaluation metric in Table 1. This creates a mild self-selection loop: CLIPScore improvements in Table 1 are not independent evidence of better text alignment. The FID results are still meaningful because FID is not part of the objective, but the text-alignment claims should be re-validated with a held-out metric or human preference data that was not used in the selection.
  4. [Appendix B / Reproducibility] Appendix B states 'we do not report error bars at this time.' The paper also does not report the optimized values of c_x and c_epsilon, the grid-search ranges, or the calibration prompt list, despite stating that the calibration set is 'hand-curated.' Several headline improvements are small in absolute FID terms (e.g., Table 1 PCM 2-step: 22.24 vs. 23.33), so without variance estimates it is hard to judge which comparisons are significant. Please add error bars or per-run results, report the selected coefficients and search ranges, and release the calibration prompts.
minor comments (5)
  1. [Section 4 (Results)] The text says Q-Sched 'outperforms PTQD noise schedules in 4/6 consistency model variants.' Inspection of Table 1 shows it outperforms PTQD in 5 of 6 variants (all except PCM at 8 steps). Please correct this count.
  2. [Table 2 caption] The caption says 'W4A8 and W8A8 are a 4x and 8x model size reduction ... yet our method improves over baseline.' This is not true for the W8A8 column, where Q-Sched is worse than both Naive and MixDQ. Please specify that the improvement statement applies to W4A8.
  3. [References] Reference [30] is a duplicate of reference [29] (Latent Consistency Models). Please merge or remove the duplicate.
  4. [Appendix H.2] The notation in Appendix H.2 defines α_t = 1 − β_t and σ_t = Π α_i, which is nonstandard and appears inconsistent with the use of α_s and α_s' in Eq. (9) as cumulative schedule quantities. Please clarify the noise schedule notation.
  5. [Figure 5b] The ablation on pre-conditioning coefficients reports PickScore, HPSV2, and JAQ Loss, but not FID. Since the paper's central claims are FID-based, reporting FID for this ablation would strengthen the analysis.

Circularity Check

2 steps flagged · score 2.0 of 10

Not circular overall: main FID claims are independent of the JAQ objective; two minor evaluation self-references (CLIPScore as both objective and metric, and JAQ-Loss row in the ablation) are non-load-bearing.

  1. fitted input called prediction [Section 3 'JAQ: A Joint Alignment Quality Loss Function' and Section 4 Table 1]
    "To learn our preconditioning coefficients, cx, cϵ, we apply grid search and evaluate each coefficient combination with respect to the JAQ loss. ... JAQ(x) = TC(x) + k · IQ(x). ... The FID and CLIPScore are calculated with respect to the COCO-30k dataset."

    In the CLIP-based variant, TC(x)=CLIPScore(x), so the two scalar coefficients are selected by maximizing CLIPScore on the calibration prompts. The paper then reports CLIPScore in Tables 1 and 2 as an evaluation result. That text-alignment evaluation is not independent: the reported CLIPScore is the same objective used for selection, so improvements on this axis are partly an artifact of the grid search. This does not affect the central FID claims because FID is not part of JAQ; indeed the W8A8 case (Q-Sched 26.34 vs Naive 25.49 on SDXL-Turbo) shows JAQ optimization does not force FID gains.

  2. fitted input called prediction [Section 4, Figure 5b ablation (Pre-Conditioning Coefficients)]
    "As shown in Figure 5b, jointly optimizing both cϵ and cx consistently yields the best results across all three metrics: PickScore, HPSv2, and JAQ Loss."

    The coefficients are selected by grid search on the JAQ loss, so the JAQ-Loss column in this ablation is the optimization objective, not an independent outcome. Reporting it as supporting evidence for joint optimization is therefore self-referential. The conclusion is not load-bearing because PickScore and HPSv2 are external to the selection objective and agree with the same conclusion.

full rationale

The main derivation is not circular. Q-Sched fits two scalar scheduler coefficients by grid search on the JAQ loss, but the headline results are measured with FID on COCO-30k/MJHQ, and FID is not a component of JAQ. Calibration prompts (hand-curated or from sDCI) are disjoint from the evaluation sets, and the method does not use full-precision inference for calibration. There are no load-bearing self-citations: the cited schedulers, PTQD, and other prior works are external. The two flagged items are evaluation-metric self-references: CLIPScore appears both in the CLIP-based JAQ objective and in the reported tables, and the ablation's JAQ-Loss row uses the selection objective as an outcome metric. Neither is load-bearing because the central FID and user-study claims are not optimized with respect to the same metrics. The user study uses actual human annotations, independent of HPSV2 used in the preference-based JAQ variant. Appendix H's 'strict guarantee' is mathematically unsupported (E[||Δx0||] is not linear in the coefficients, and the zero-error minimum is unattainable with fixed quantization error), but this is a correctness defect, not a circularity. Likewise, the W8A8 FID regression on SDXL-Turbo (26.34 vs Naive 25.49) contradicts the unqualified abstract claim but demonstrates that JAQ selection does not force FID outcomes, i.e., the FID evaluation is not circular. Overall the method is self-contained against external FID and human-preference benchmarks, so the circularity score is low.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The ledger shows Q-Sched is cheap in terms of free parameters (two grid-searched scalars and one hand-tuned loss weight) but relies on strong domain assumptions about sampler correctness, coefficient sufficiency, metric validity, and calibration generalization. No invented entities are introduced.

free parameters (3)
  • cx (sample-side preconditioning coefficient) = not reported
    Learned by grid search with JAQ loss on 5-20 calibration prompts; exact values are never reported in the manuscript, preventing replication.
  • c_epsilon (model-side preconditioning coefficient) = not reported
    Same as cx; chosen jointly with cx by grid search, but the grid ranges and selected values are absent.
  • k (JAQ loss trade-off) = 2
    Hand-tuned; the paper acknowledges a hand-tuned k is sufficient and uses k = 2 for all experiments, with an ablation showing sensitivity at k = 5.
assumptions (4)
  • domain assumption TCD strategic stochastic sampling formula (Eq. 9) correctly describes the few-step diffusion process for LCM/PCM variants.
    Q-Sched modifies Eq. 9; if the underlying sampler model is wrong for a given backbone, the two-coefficient correction may not transfer.
  • domain assumption A single pair of scalar coefficients applied uniformly at every denoising step can correct the trajectory shift caused by quantization.
    This is the paper's load-bearing modeling choice, stated in Section 3 ('We find that two coefficients are sufficient'). No proof is given that one pair works across models, bit-widths, and prompts.
  • domain assumption CLIPScore and CLIP-IQA (or AQ-MAP and HPSV2) are valid proxies for the image quality and text alignment of quantized outputs.
    The JAQ loss is built from these metrics; if these metrics are insensitive to quantization artifacts, the selected coefficients may not reflect true quality.
  • domain assumption Hand-curated 5-20 prompt calibration sets are representative enough to select coefficients that generalize to COCO-30k and MJHQ-30k.
    The paper uses a compact calibration set and reuses it across evaluations; generalization of grid-selected coefficients depends on this representativeness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Q-Sched: Pushing the Boundaries of Few-Step Diffusion Models with Quantization-Aware Scheduling." pith.science (2026). https://pith.science/paper/QEYLI6V6

@misc{pith2026250901624,
  author       = {Pith},
  title        = {Pith review of: Q-Sched: Pushing the Boundaries of Few-Step Diffusion Models with Quantization-Aware Scheduling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QEYLI6V6}},
  note         = {Machine review of arXiv:2509.01624}
}
read the original abstract

Text-to-image diffusion models are computationally intensive, often requiring dozens of forward passes through large transformer backbones. For instance, Stable Diffusion XL generates high-quality images with 50 evaluations of a 2.6B-parameter model, an expensive process even for a single batch. Few-step diffusion models reduce this cost to 2-8 denoising steps but still depend on large, uncompressed U-Net or diffusion transformer backbones, which are often too costly for full-precision inference without datacenter GPUs. These requirements also limit existing post-training quantization methods that rely on full-precision calibration. We introduce Q-Sched, a new paradigm for post-training quantization that modifies the diffusion model scheduler rather than model weights. By adjusting the few-step sampling trajectory, Q-Sched achieves full-precision accuracy with a 4x reduction in model size. To learn quantization-aware pre-conditioning coefficients, we propose the JAQ loss, which combines text-image compatibility with an image quality metric for fine-grained optimization. JAQ is reference-free and requires only a handful of calibration prompts, avoiding full-precision inference during calibration. Q-Sched delivers substantial gains: a 15.5% FID improvement over the FP16 4-step Latent Consistency Model and a 16.6% improvement over the FP16 8-step Phased Consistency Model, showing that quantization and few-step distillation are complementary for high-fidelity generation. A large-scale user study with more than 80,000 annotations further confirms Q-Sched's effectiveness on both FLUX.1[schnell] and SDXL-Turbo.

Figures

Figures reproduced from arXiv: 2509.01624 by the authors.

Figure 1
Figure 1. Q-Sched introduces a quantization-aware noise scheduler to few-step diffusion backbones and achieves excellent image fidelity. We find quantization and few-step diffusions to be complementary model compression strategies. Abstract Text-to-image diffusion models are computationally intensive, often requiring dozens of forward passes through large transformer backbones. For instance, Stable Diffusion XL generates high… view at source ↗
Figure 2
Figure 2. Q-Sched’s quantization-aware scheduling enables state-of-the-art image fidelity across multiple compressed few-step models. Q-Sched directly optimizes the few-step diffusion’s scheduler (see Figure 2b) whereas prior work directly optimizes the transformer or U-Net backbone. we propose the JAQ loss, which combines text-image compatibility with an image quality metric for fine-grained optimization. JAQ is reference-fr… view at source ↗
Figure 3
Figure 3. 4-Step (top row) and 8-Step (bottom row) LCMs. Prompt: "a car and a bus on a french [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Comparing Q-Sched across various bit widths. Despite this, a small-scale user study with 52 anonymized par￾ticipants showed that Q-Sched was preferred by 56% of users over MixDQ in the W8A8 setting, indicating that perceptual quality may remain competitive even when FI…
Figure 5
Figure 5. Figure 5: Ablation studies on various design choices for [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Optimizing Q-Sched with various reference-less image quality metrics. Our loss function, JAQ, is a linear combination of CLIPScore and CLIP-IQA-Q. We compare against three CLIP-IQA prompts: Complexity, Noisiness, and Quality denoted as -C, -N, -Q respectively [PITH_FU…
Figure 7
Figure 7. Figure 7: Three types of image artifacts that occur when quantizing image generation models. Images [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: FID on COCO-30k. A 4W8A compressed model with our [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: Selected prompts where Q-Sched outperforms full precision. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]
Figure 10
Figure 10. Figure 10: Selected prompts where Q-Sched fails to improve over 4W8A. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_10.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Quantizing Recursive Reasoning Models

    cs.LG 2026-06 conditional novelty 6.0 of 10

    Per-tensor 4-bit quantization of weight-tied recursive reasoners induces accumulating activation bias that kills accuracy; per-block scaling (MXInt4) restores it.

  2. Holding the FP8 Quality Ceiling at 8-Bit Weights and Activations: INT8 and GGUF Post-Training Quantization of Ideogram 4.0 for Consumer GPUs

    cs.LG 2026-06 unverdicted novelty 4.0 of 10

    INT8 W8A8 post-training quantization of Ideogram 4.0 preserves FP8 quality on a 200-prompt benchmark while outperforming NF4 on CLIP score and offering a favorable quality-memory trade-off via GGUF Q4_K.

Reference graph

Works this paper leans on

55 extracted references · 27 canonical work pages · cited by 2 Pith papers

  1. [1]

    URL https://www.rapidata.ai/

    Rapidata: An api that provides fast access to large-scale human evaluations, 2025. URL https://www.rapidata.ai/. Accessed: 2025-05-16

    work page 2025

  2. [2]

    Segdiff: Image segmentation with diffusion probabilistic models

    Tomer Amit, Eliya Nachmani, Tal Shaharbany, and Lior Wolf. Segdiff: Image segmentation with diffusion probabilistic models. arXiv preprint arXiv:2112.00390, 2021

    arXiv 2021

  3. [3]

    Protein structure and sequence generation with equivariant denoising diffusion probabilistic models

    Namrata Anand and Tudor Achim. Protein structure and sequence generation with equivariant denoising diffusion probabilistic models. arXiv preprint arXiv:2205.15019, 2022

    arXiv 2022

  4. [4]

    Structured denoising diffusion models in discrete state-spaces

    Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 34:17981–17993, 2021

    work page 2021

  5. [5]

    Blended diffusion for text-driven editing of natural images

    Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208–18218, 2022

    work page 2022

  6. [6]

    Label-efficient semantic segmentation with diffusion models

    Dmitry Baranchuk, Ivan Rubachev, Andrey V oynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021

    arXiv 2021

  7. [7]

    Flux.1-schnell

    Black Forest Labs. Flux.1-schnell. https://huggingface.co/black-forest-labs/FLUX. 1-schnell, 2024. Accessed: 2025-05-14

    work page 2024

  8. [8]

    Denoising pretraining for semantic segmentation

    Emmanuel Asiedu Brempong, Simon Kornblith, Ting Chen, Niki Parmar, Matthias Minderer, and Mohammad Norouzi. Denoising pretraining for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4175–4186, 2022

    work page 2022

Show all 55 references
  1. [9]

    High-frequency space diffusion models for accelerated mri

    Chentao Cao, Zhuo-Xu Cui, Shaonan Liu, Dong Liang, and Yanjie Zhu. High-frequency space diffusion models for accelerated mri. arXiv preprint arXiv:2208.05481, 2022

    2022 arXiv

  2. [10]

    Q-dit: Accurate post-training quantization for diffusion transformers

    Lei Chen, Yuan Meng, Chen Tang, Xinzhu Ma, Jingyan Jiang, Xin Wang, Zhi Wang, and Wenwu Zhu. Q-dit: Accurate post-training quantization for diffusion transformers. arXiv preprint arXiv:2406.17343, 2024

    2024 arXiv

  3. [11]

    Structural pruning for diffusion models

    Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. Advances in neural information processing systems, 36, 2024

    2024

  4. [12]

    Squant: On-the-fly data-free quantization via diagonal hessian approximation

    Cong Guo, Yuxian Qiu, Jingwen Leng, Xiaotian Gao, Chen Zhang, Yunxin Liu, Fan Yang, Yuhao Zhu, and Minyi Guo. Squant: On-the-fly data-free quantization via diagonal hessian approximation. arXiv preprint arXiv:2202.07471, 2022

    2022 arXiv

  5. [13]

    Ptqd: Accurate post-training quantization for diffusion models

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

    2024

  6. [14]

    Clipscore: A reference-free evaluation metric for image captioning

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

    2021 arXiv

  7. [15]

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

    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 equilibrium. Advances in neural information processing systems, 30, 2017

    2017

  8. [16]

    Denoising diffusion probabilistic models

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

    2020

  9. [17]

    Cascaded diffusion models for high fidelity image generation

    Jonathan Ho, Chitwan Saharia, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 23(47):1–33, 2022. 11

    2022

  10. [18]

    Knowledge diffusion for distillation

    Tao Huang, Yuan Zhang, Mingkai Zheng, Shan You, Fei Wang, Chen Qian, and Chang Xu. Knowledge diffusion for distillation. Advances in Neural Information Processing Systems, 36, 2024

    2024

  11. [19]

    Rethinking fid: Towards a better evaluation metric for image generation

    Sadeep Jayasumana, Srikumar Ramalingam, Andreas Veit, Daniel Glasner, Ayan Chakrabarti, and Sanjiv Kumar. Rethinking fid: Towards a better evaluation metric for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9307–9...

    2024

  12. [20]

    Elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364, 2022

    2022 arXiv

  13. [21]

    Q-refine: A perceptual quality refiner for ai-generated image

    Chunyi Li, Haoning Wu, Zicheng Zhang, Hongkun Hao, Kaiwei Zhang, Lei Bai, Xiaohong Liu, Xiongkuo Min, Weisi Lin, and Guangtao Zhai. Q-refine: A perceptual quality refiner for ai-generated image. In 2024 IEEE International Conference on Multimedia and Expo (ICME), pages 1–6. IEEE, 2024

    2024

  14. [22]

    Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation, 2024

    Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation, 2024

    2024

  15. [23]

    Svdquant: Absorbing outliers by low-rank component for 4-bit diffusion models

    Muyang Li, Yujun Lin, Zhekai Zhang, Tianle Cai, Junxian Guo, Xiuyu Li, Enze Xie, Chenlin Meng, Jun-Yan Zhu, and Song Han. Svdquant: Absorbing outliers by low-rank component for 4-bit diffusion models. In The Thirteenth International Conference on Learning Representations, 2025

    2025

  16. [24]

    Diffusion-lm improves controllable text generation

    Xiang Li, John Thickstun, Ishaan Gulrajani, Percy S Liang, and Tatsunori B Hashimoto. Diffusion-lm improves controllable text generation. Advances in Neural Information Processing Systems, 35:4328–4343, 2022

    2022

  17. [25]

    Q-diffusion: Quantizing diffusion models

    Xiuyu Li, Long Lian, Yijiang Liu, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. arXiv preprint arXiv:2302.04304, 2023

    2023 arXiv

  18. [26]

    Q-dm: An efficient low-bit quantized diffusion model

    Yanjing Li, Sheng Xu, Xianbin Cao, Xiao Sun, and Baochang Zhang. Q-dm: An efficient low-bit quantized diffusion model. Advances in Neural Information Processing Systems, 36, 2024

    2024

  19. [27]

    Microsoft coco: Common objects in context

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

    2014

  20. [28]

    Instaflow: One step is enough for high-quality diffusion-based text-to-image generation

    Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In The Twelfth International Conference on Learning Representations, 2023

    2023

  21. [30]

    Latent consistency models: Synthesizing high-resolution images with few-step inference

    Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023

    2023 arXiv

  22. [31]

    Sdedit: Guided image synthesis and editing with stochastic differential equations

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2021

    2021

  23. [32]

    No-reference image quality assessment in the spatial domain

    Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing, 21(12):4695–4708, 2012

    2012

  24. [33]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023. 12

    2023

  25. [34]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023

    2023 arXiv

  26. [35]

    Hierarchical text-conditional image generation with clip latents

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022

    2022 arXiv

  27. [36]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684–10695, 2022

    2022

  28. [37]

    U-net: Convolutional networks for biomedical image segmentation

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

    2015

  29. [38]

    Adversarial diffusion distillation

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87–103. Springer, 2024

    2024

  30. [39]

    Post-training quantization on diffusion models

    Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. arXiv preprint arXiv:2211.15736, 2022

    2022 arXiv

  31. [40]

    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, 36, 2024

    2024

  32. [41]

    Denoising diffusion implicit models

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

    2010 arXiv

  33. [42]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations, 2021

    2021

  34. [43]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023

    2023 arXiv

  35. [44]

    Towards accurate data-free quantization for diffusion models

    Changyuan Wang, Ziwei Wang, Xiuwei Xu, Yansong Tang, Jie Zhou, and Jiwen Lu. Towards accurate data-free quantization for diffusion models. arXiv preprint arXiv:2305.18723, 2(5), 2023

    2023 arXiv

  36. [45]

    Phased consistency model

    Fu-Yun Wang, Zhaoyang Huang, Alexander William Bergman, Dazhong Shen, Peng Gao, Michael Lingelbach, Keqiang Sun, Weikang Bian, Guanglu Song, Yu Liu, et al. Phased consistency model. arXiv preprint arXiv:2405.18407, 2024

    2024 arXiv

  37. [46]

    Quest: Low-bit diffusion model quantization via efficient selective finetuning

    Haoxuan Wang, Yuzhang Shang, Zhihang Yuan, Junyi Wu, and Yan Yan. Quest: Low-bit diffusion model quantization via efficient selective finetuning. arXiv preprint arXiv:2402.03666, 2024

    2024 arXiv

  38. [47]

    Exploring clip for assessing the look and feel of images

    Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2555–2563, 2023

    2023

  39. [48]

    Image quality assessment: from error visibility to structural similarity

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

    2004

  40. [49]

    Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis

    Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. CoRR, 2023

    2023

  41. [50]

    Diffusion probabilistic modeling for video generation

    Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion probabilistic modeling for video generation. arXiv preprint arXiv:2203.09481, 2022

    2022 arXiv

  42. [51]

    Cogvideox: Text-to-video diffusion models with an expert transformer

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 13

    2024 arXiv

  43. [52]

    One-step diffusion with distribution matching distillation

    Tianwei Yin, Michaël Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. arXiv preprint arXiv:2311.18828, 2023

    2023 arXiv

  44. [53]

    Mixdq: Memory-efficient few-step text-to-image diffusion models with metric-decoupled mixed precision quantization

    Tianchen Zhao, Xuefei Ning, Tongcheng Fang, Enshu Liu, Guyue Huang, Zinan Lin, Shengen Yan, Guohao Dai, and Yu Wang. Mixdq: Memory-efficient few-step text-to-image diffusion models with metric-decoupled mixed precision quantization. In European Conference on Computer Vision, p...

    2024

  45. [54]

    Which image is of higher quality and more aesthetically pleasing?

    Jianbin Zheng, Minghui Hu, Zhongyi Fan, Chaoyue Wang, Changxing Ding, Dacheng Tao, and Tat-Jen Cham. Trajectory consistency distillation. arXiv preprint arXiv:2402.19159, 2024. 14 A Details on User Preference Assessment We design our evaluation setup following the user prefere...

    2024 arXiv

  46. [55]

    The quantized model’s distribution shift can be modeled through a linear correction term

  47. [56]

    Why is Q-Sched able to learn a better noise schedule when it is also a linear correction?

    The uncorrelated quantization noise is normally distributed. While these assumptions are similar to prior work on diffusion models, they are likely to break down on the few-step diffusions where the denoising process is distilled from many steps and is not expected to be linea...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.