Pith. sign in

REVIEW 4 major objections 6 minor 63 references

Text Embedding Knows How to Quantize Text-Guided Diffusion Models

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that text prompt embeddings can predict how much quantization a diffusion model tolerates, enabling per-layer, per-timestep bit precision selection that cuts cost without hurting image quality.

desk verdict Clean, well-motivated quantization wrapper with plausible gains, but the missing no-text control means the headline 'text guides bits' claim is under-supported. read the letter →

arxiv 2507.10340 v3 pith:C6HIKF6S submitted 2025-07-14 cs.CV

Hongjae Lee , Myungjun Son , Dongjea Kang , Seung-Won Jung This is my paper · ORCID
classification cs.CV
keywords text-guideddiffusionmodelspost-trainingquantizationdynamicbitprecisiontext-promptconditioningimagequalitypredictionCLIPembeddingscomputationalefficiency
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

QLIP claims that the text prompt given to a text-to-image diffusion model is a reliable guide for deciding how aggressively each layer can be quantized at each denoising step. The paper trains a small text-to-quality predictor on CLIP embeddings and images generated by the full-precision model, then uses the predicted quality to assign low, medium, or high bit precision per layer per time step. Wrapped around Q-diffusion or PTQD, QLIP lowers the average activation bit width while keeping or improving FID and sFID on BK-SDM-Tiny-2M, Stable Diffusion v1.4, and SDXL across COCO2017 and Conceptual Captions. The consequence is that a prompt-derived quality signal can make quantization cost-adaptive without sacrificing output quality.

What carries the argument

The load-bearing mechanism is a two-stage mapping from text to bits. A text-to-quality (T2Q) module, a three-layer MLP on top of the CLIP text embedding, predicts the score $q = \phi(z)$ that a full-precision generated image would receive under the GIQA quality metric. A quality-to-bit (Q2B) module then forms selection probabilities $p_{\text{low}}, p_{\text{med}}, p_{\text{high}}$ from $q$ and from learnable per-time-step parameters, takes the argmax as the bit width for each layer, and makes the discrete choice differentiable with a straight-through estimator so that Q2B can be trained to minimize both the output error against the full-precision denoiser and a bit-cost regularizer.

What would settle it

Measure the Spearman correlation between the frozen T2Q predictions and actual GIQA scores on images generated by BK-SDM or SDXL from Conceptual Captions prompts; if that correlation is much below the 0.8047 reported for Stable Diffusion v1.4 on COCO, the prompt-to-quality signal no longer supports QLIP's bit allocation and the reported FAB and FID gains should not transfer.

Watch

Extended reading notes

Core claim

The central claim is that the image quality a text-guided diffusion model will produce is predictable from the prompt's CLIP text embedding, and this prediction is enough to choose where bits should be spent during quantization. QLIP implements this with a frozen text-to-quality (T2Q) module that maps the embedding to a normalized quality score, and a trained quality-to-bit (Q2B) module that converts the score plus learnable time-step parameters into per-layer, per-timestep bit selections among three precisions. The paper reports that applying QLIP to Q-diffusion and PTQD reduces feature average bit width (for example from 16 to 12.14 on BK-SDM-Tiny-2M and from 8 to 7.86 on Stable Diffusion v1.4) while matching or improving FID and sFID, and that the benefit persists on SDXL and on the mixed-precision baseline PCR.

Load-bearing premise

The paper's load-bearing premise is that one frozen text-to-quality predictor, trained only on Stable Diffusion v1.4 images generated from COCO captions, continues to predict image quality accurately enough to assign bit widths for other diffusion models and other caption sources.

Editorial extensions

If this is right

  • Adding QLIP to Q-diffusion or PTQD lowers average activation bit width by roughly a quarter with FID and sFID maintained or improved.
  • Detailed prompts automatically cost more compute than terse prompts, because high predicted quality routes high and medium bits to most layers and time steps.
  • The quality signal transfers beyond its training setup: QLIP reduces FAB on SDXL and on the FLUX-based mixed-precision baseline PCR.
  • Bit savings concentrate in cross-attention layers (proj in, at2.to v, proj out), so those blocks are the main low-bit candidates.

Reading between the lines

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

  • The paper never re-measures text-to-quality accuracy after transfer; checking SROCC on SDXL and BK-SDM with Conceptual Captions would show whether the frozen predictor truly generalizes.
  • If prompt-to-quality mapping drives the gain, cheaper input features such as caption length or a small reward model could substitute for the trained T2Q; the paper's ablations already show predicted quality beats prompt length and image complexity.
  • Because bit width is taken as the maximum across a batch, QLIP's savings shrink as batch size grows, pointing to per-request or small-batch deployment as the practical setting.
  • A natural extension would let the same quality signal also choose the number of denoising steps or the sampler, making the entire sampling budget prompt-adaptive rather than only the bit widths.
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 QLIP, a post-training quantization method for text-guided diffusion models that selects per-layer, per-time-step activation bit precisions using the input text prompt. QLIP consists of two modules: a Text-to-Quality (T2Q) module, a three-layer MLP trained with Eq. (3) to predict the GIQA score of full-precision generated images from the CLIP text embedding, and a Quality-to-Bit (Q2B) module, which turns the predicted quality q into bit-selection probabilities via Eqs. (4)-(6) and a straight-through estimator (Eq. (8)). The Q2B module is trained with the loss in Eq. (9), combining quantization error with a bit-length penalty. The method is integrated into Q-diffusion and PTQD and evaluated on BK-SDM-Tiny-2M, Stable Diffusion v1.4, and SDXL on COCO2017 and Conceptual Captions, with additional results on FLUX with PCR in the supplementary. The main reported findings are reductions in average bit width (FAB) and BitOPs while maintaining or improving FID/sFID relative to fixed-bit baselines, plus runtime and ablation studies.

Significance. If the central claim holds, text-derived quality is a novel and transferable signal for cost-effective dynamic quantization of diffusion models, and the paper would be a useful contribution to efficient text-to-image generation. The manuscript has several strengths: the pipeline is clearly specified; the code repository is provided; the experiments span multiple base models (BK-SDM-Tiny-2M, SD v1.4, SDXL) and datasets; the supplementary includes a comparison with a recent mixed-precision method (PCR) on FLUX; and the runtime and memory overhead of the added modules are measured. The motivating observation in Figure 2, that quantization sensitivity correlates with full-precision image quality, is interesting and well illustrated. However, the causal role of the text embedding is not yet isolated, and the transfer validity of the T2Q module to models and datasets outside its training distribution is unmeasured. These gaps are central to the paper's headline claim and need to be addressed before the results can be fully trusted.

major comments (4)
  1. [Section 3.4, Eq. (5) and Tables 1-2, S1-S2] The paper does not include a no-text control, so the central claim that text prompts guide bit precision is not isolated. In Eq. (5), u_t_m and u_t_h are prompt-independent learnable parameters that alone define a per-layer, per-time-step mixed-precision policy, while pq in Eq. (4) is the only text-dependent term. All comparisons in Tables 1, 2, S1, and S2 are against fixed-bit baselines (Q-diffusion W4A16/W4A8, PTQD, PCR), not against a static or learned mixed-precision policy with pq removed. Table 5 ablates pq only in combination with pt_m/pt_h, and Table S3 compares alternative input signals (image complexity, prompt length) but not a text-free Q2B. The reported FAB reductions and FID improvements could therefore be entirely due to the learned time/layer bit allocation, which is not new (PTQD and TDQ already exploit time-step sensitivity, and per-layer mixed precision is standard in quantization literature). Please add an ablation that replaces pq with a constant (or removes the text branch) while keeping the Q2B training procedure and bit options otherwise identical.
  2. [Section 4.1 and Table 4] The T2Q module is trained on 10k COCO2017 captions with Stable Diffusion v1.4, and its predictive accuracy is measured only on the same model and dataset (SROCC 0.8047 in Table 4). The frozen T2Q is then deployed on BK-SDM-Tiny-2M, SDXL, and Conceptual Captions without reporting per-model or per-dataset T2Q accuracy. If the text-to-quality correlation weakens on these targets, QLIP would systematically misallocate bits, undercutting the reported gains. Please report SROCC/PLCC of the frozen T2Q on each target model/dataset, or alternatively provide an analysis that groups evaluation prompts by predicted quality and shows that the bit-allocation pattern remains sensible (e.g., FID per quality bin).
  3. [Tables 1, 2, S1, S2] All quantitative comparisons lack error bars, confidence intervals, or significance tests. Several claimed 'maintained' or 'improved' results are within typical sampling noise: for example, in Table 1 (COCO2017), Q-diffusion FID is 30.02 and +QLIP is 30.01, and sFID is 73.25 vs. 73.24; in Table 2 (Conceptual Captions), the FID improvement of 20.74 to 19.89 is reported without variance. Please run multiple seeds or multiple generation sets for at least the headline comparisons (Tables 1 and 2) and report mean and standard deviation, or otherwise justify that the observed differences are significant.
  4. [Section 3.1 and Section B (supplementary)] The paper states in Section 3.1 that 'To our knowledge, all previous diffusion quantization methods set the same bit precision to all layers,' but the supplementary (Section B) compares against PCR [54], which is described there as 'a recent mixed-precision quantization method.' This is internally inconsistent: if PCR already performs per-layer mixed-precision quantization, the novelty claim in Section 3.1 is too strong and should be revised. Please clarify whether PCR assigns different bit precisions across layers and, if so, reposition QLIP's contribution accordingly (e.g., as text-conditional per-layer mixed precision rather than the first per-layer mixed-precision method).
minor comments (6)
  1. [Section 3.1, Eq. (1)] There is a typo in Eq. (1): 'parameteres' should be 'parameters'.
  2. [Section 4.1, Hyperparameter settings] The hyperparameters m (number of initial high-bit time steps) and the PTQD noise-correction thresholds 0.3/0.7 are hand-set and not ablated or justified. Figure 4 ablates only lambda_bit and M. Please add a sensitivity study or at least a brief explanation for these choices.
  3. [Table S3] The prompt-length baseline in Table S3 is a simple heuristic and not a trained alternative, and the image-complexity baseline changes the input signal but keeps the Q2B architecture. For a fair comparison, use a trained text-free Q2B with the same bit options and training budget, which would also serve as the no-text control requested in the major comments.
  4. [Section 4.2, Table 3] The runtime measurement in Table 3 does not describe how mixed-precision per-layer execution is realized with Cutlass, nor whether the reported times include per-layer kernel-launch overhead or the cost of the T2Q/Q2B modules in a batched setting. Please clarify the measurement protocol.
  5. [Tables 1 and 2] The bit-option sets differ across experiments: W4A{8,16,32} in Table 1 and W4A{6,8,10} in Table 2. Table S4 provides some analysis, but only for one setting. Please justify the choice of bit options per model or report results for a common set.
  6. [Section 4.4, Table 5] The Q2B ablations report FAB and FID separately without discussing the trade-off at matched operating points. For instance, '+pq+pt_h' achieves FAB 6.73 but FID 29.37, which is worse than the Q-diffusion W4A8 baseline (FID 23.40). Please analyze the results at comparable FAB or comparable FID levels.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: QLIP's bit allocation is learned from quantization error plus a bit penalty, and its text-quality predictor is trained on full-precision GIQA targets, so the reported FID/FAB gains are not forced by construction.

full rationale

The central claim is an empirical training pipeline, not a derivation that reduces to its inputs. The T2Q module is trained with Eq. (3), where the targets are GIQA scores of images generated by the full-precision diffusion model; these targets are independent of the quantized metrics (FID/sFID) reported later. The Q2B module is trained with Eq. (9), which combines the full-precision-versus-quantized prediction error with an explicit bit-length penalty, and the final bit precision in Eq. (7) is an argmax over learned probabilities that are optimized rather than fitted to the reported FID. The main comparisons (Tables 1, 2, S1, S2) are against fixed-bit and existing mixed-precision baselines on external validation sets, and the evaluation metrics FID, sFID, and CLIP Score are external to the training objectives. The self-citation to RefQSR [25] appears only as a related-work example of image-adaptive quantization and as one of two sources for the BitOPs metric; it is not load-bearing, and no uniqueness theorem or ansatz is imported from it. The absence of a text-free Q2B control is a genuine experimental-isolation concern: the observed bit savings could in principle be driven largely by the learned per-layer, per-time-step offsets u_t_m and u_t_h rather than by the text-dependent quality signal. However, that is an experimental gap, not circularity, because no equation in the paper is equivalent to its own inputs by construction and no reported quantity is a renamed fitted parameter.

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

The central claim rests on a chain of empirical assumptions: the text-to-quality correlation (validated only on SD v1.4/COCO), the activation-distribution grouping by time steps, the fixed 4-bit weights, and the hand-set thresholds/hyperparameters. All are either standard practice or domain assumptions from prior work; none is a new physical entity.

free parameters (4)
  • lambda_bit = 1
    Weighting between quantization error and bit-length penalty in Eq. 9; selected on COCO validation via Figure 4(a); all main tables use this value.
  • M (time-step grouping) = 200
    Grouping size for time-step parameters u_t_m and u_t_h; chosen via Figure 4(b) on COCO validation; determines temporal granularity of bit assignment.
  • m (initial high-bit time steps) = not reported
    Number of first denoising steps where p_q is forced to 1, so high bits are always selected (Section 3.4); value is never given, yet it shapes early denoising and likely affects FID.
  • PTQD noise-correction thresholds = 0.3 and 0.7
    Noise correction is disabled when predicted quality is below 0.3 or above 0.7 (Section 4.1); manual thresholds used for all PTQD+QLIP results, not ablated.
assumptions (6)
  • domain assumption Image quality of the generated image is predictable from the CLIP text embedding.
    Used in Section 3.3; the T2Q module is trained to regress GIQA from CLIP embeddings; validation on COCO gives SROCC 0.8047 (Table 4), but the transfer to other models/datasets is assumed.
  • domain assumption Low-bit quantization is acceptable when the full-precision output would be low quality; high bits are needed when output quality is high.
    Motivating observation in Figure 2; underlies the whole quality-to-bit mapping, but Figure 2 shows no quantitative correlation.
  • domain assumption Activation distributions vary across time steps and layers but are similar for adjacent time steps, allowing parameter sharing every M steps.
    Invoked in Section 3.4, citing [15,50]; justifies u^t_j shared across M-step groups.
  • domain assumption The initial stage of reverse denoising needs high bits to maintain text-image correlation.
    Section 3.4, based on [3,8]; used to force p_q=1 for the first m time steps.
  • domain assumption Weights are less sensitive to quantization than activations for diffusion models.
    Section 3.4, citing [15,27]; allows fixing weights at 4 bits for all layers and time steps.
  • standard math Straight-through estimator gives usable gradients for the non-differentiable bit-selection argmax.
    Eq. 8; standard practice, but known to be a biased gradient estimator.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Text Embedding Knows How to Quantize Text-Guided Diffusion Models." pith.science (2026). https://pith.science/paper/C6HIKF6S

@misc{pith2026250710340,
  author       = {Pith},
  title        = {Pith review of: Text Embedding Knows How to Quantize Text-Guided Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C6HIKF6S}},
  note         = {Machine review of arXiv:2507.10340}
}
read the original abstract

Despite the success of diffusion models in image generation tasks such as text-to-image, the enormous computational complexity of diffusion models limits their use in resource-constrained environments. To address this, network quantization has emerged as a promising solution for designing efficient diffusion models. However, existing diffusion model quantization methods do not consider input conditions, such as text prompts, as an essential source of information for quantization. In this paper, we propose a novel quantization method dubbed Quantization of Language-to-Image diffusion models using text Prompts (QLIP). QLIP leverages text prompts to guide the selection of bit precision for every layer at each time step. In addition, QLIP can be seamlessly integrated into existing quantization methods to enhance quantization efficiency. Our extensive experiments demonstrate the effectiveness of QLIP in reducing computational complexity and improving the quality of the generated images across various datasets.

Figures

Figures reproduced from arXiv: 2507.10340 by the authors.

Figure 1
Figure 1. Examples of the images generated by Stable Diffusion v1.4 using the text prompts given below images. When the text prompt [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Analysis of the effect of quantization and the quality of generated images, where the image quality is measured using the GIQA [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of QLIP that consists of two key modules: T2Q module that predicts image quality from the input text prompt and [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Results on Q-diffusion using different hyperparameter [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Examples of variations in FAB by QLIP for the texts with different levels of richness and detail, along with the generated images. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Examples of the bit selection results and generated images. When the text prompt is specifically describing the image to be [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

63 extracted references · 59 canonical work pages

  1. [54]

    Post-training quan- tization with progressive calibration and activation relaxing for text-to-image diffusion models

    Siao Tang, Xin Wang, Hong Chen, Chaoyu Guan, Zewen Wu, Yansong Tang, and Wenwu Zhu. Post-training quan- tization with progressive calibration and activation relaxing for text-to-image diffusion models. In Proceedings of the European Conference on Computer Vision, pages 404–420,

    work page

  2. [1]

    FLUX: huggingface.co/black-forest-labs/flux.1-dev. 1

    work page

  3. [2]

    GPT Image 1 API: https://platform.openai.com/docs/models/gpt- image-1. 3

    work page

  4. [3]

    eDiff-I: Text-to-image diffusion models with ensemble of expert denoisers

    Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Ait- tala, Timo Aila, Samuli Laine, et al. eDiff-I: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 5

    arXiv 2022

  5. [4]

    Analytic- DPM: An analytic estimate of the optimal reverse variance in diffusion probabilistic models

    Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic- DPM: An analytic estimate of the optimal reverse variance in diffusion probabilistic models. In Proceedings of the In- ternational Conference on Learning Representations , 2022. 1

    work page 2022

  6. [5]

    Estimating or propagating gradients through stochastic neurons for conditional computation

    Yoshua Bengio, Nicholas L ´eonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 5

    arXiv 2013

  7. [6]

    Improving image generation with better captions

    James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn. openai. com/papers/dall-e-3. pdf, 2023. 1

    work page 2023

  8. [7]

    PACT: Parameterized clipping acti- vation for quantized neural networks

    Jungwook Choi, Zhuo Wang, Swagath Venkataramani, Pierce I-Jen Chuang, Vijayalakshmi Srinivasan, and Kailash Gopalakrishnan. PACT: Parameterized clipping acti- vation for quantized neural networks. arXiv preprint arXiv:1805.06085, 2018. 2

    arXiv 2018

Show all 63 references
  1. [8]

    Perception pri- oritized training of diffusion models

    Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception pri- oritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472–11481, 2022. 5

    2022

  2. [9]

    QNCD: Quantization noise correction for diffusion models

    Huanpeng Chu, Wei Wu, Chengjie Zang, and Kun Yuan. QNCD: Quantization noise correction for diffusion models. arXiv preprint arXiv:2403.19140, 2024. 2

    2024 arXiv

  3. [10]

    Mix- Path: A unified approach for one-shot neural architecture search

    Xiangxiang Chu, Shun Lu, Xudong Li, and Bo Zhang. Mix- Path: A unified approach for one-shot neural architecture search. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 5972–5981, 2023. 1

    2023

  4. [11]

    HAWQ: Hessian aware quantiza- tion of neural networks with mixed-precision

    Zhen Dong, Zhewei Yao, Amir Gholami, Michael W Ma- honey, and Kurt Keutzer. HAWQ: Hessian aware quantiza- tion of neural networks with mixed-precision. In Proceed- ings of the IEEE/CVF International Conference on Com- puter Vision, pages 293–302, 2019. 1, 3

    2019

  5. [12]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Proceedings of the Advances in Neural Information Processing Systems , pages 2672–2680, 2014. 1

    2014

  6. [13]

    GIQA: Generated image quality assessment

    Shuyang Gu, Jianmin Bao, Dong Chen, and Fang Wen. GIQA: Generated image quality assessment. In Proceedings of the European Conference on Computer Vision, pages 369– 385, 2020. 3, 4, 8

    2020

  7. [14]

    Learning both weights and connections for efficient neural network

    Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. In Proceedings of the Advances in Neural Information Process- ing Systems, pages 1135—-1143, 2015. 1

    2015

  8. [15]

    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. InProceedings of the Advances in Neu- ral Information Processing Systems , pages 13237–13249,

  9. [16]

    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. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Pro- cessing, pages 7514–7528, 2021. 5

    2021

  10. [17]

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

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilib- rium. In Proceedings of the Advances in Neural Information Processing Systems, pages 6629—-6640, 2017. 5

    2017

  11. [18]

    Denoising diffu- sion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models. In Proceedings of the Advances in Neural Information Processing Systems , pages 6840–6851,

  12. [19]

    AdaBM: On-the-fly adaptive bit mapping for image super-resolution

    Cheeun Hong and Kyoung Mu Lee. AdaBM: On-the-fly adaptive bit mapping for image super-resolution. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2641–2650, 2024. 1, 3, 6

    2024

  13. [20]

    CADyQ: Content-aware dynamic quantization for image super-resolution

    Cheeun Hong, Sungyong Baik, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. CADyQ: Content-aware dynamic quantization for image super-resolution. In Proceedings of the European Conference on Computer Vision , pages 367– 383, 2022. 1, 2, 3, 6

    2022

  14. [21]

    BK-SDM: A lightweight, fast, and cheap version of stable diffusion

    Bo-Kyeong Kim, Hyoung-Kyu Song, Thibault Castells, and Shinkook Choi. BK-SDM: A lightweight, fast, and cheap version of stable diffusion. In Proceedings of the European Conference on Computer Vision, 2024. 2, 5, 6

    2024

  15. [22]

    Adam: A method for stochastic optimization

    Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations, 2015. 5

    2015

  16. [23]

    Pick-a-Pic: An open dataset of user preferences for text-to-image generation

    Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Ma- tiana, Joe Penna, and Omer Levy. Pick-a-Pic: An open dataset of user preferences for text-to-image generation. In Proceedings of the Advances in Neural Information Process- ing Systems Workshop, 2023. 1

    2023

  17. [24]

    Improved precision and recall met- ric for assessing generative models

    Tuomas Kynk ¨a¨anniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall met- ric for assessing generative models. In Proceedings of the Advances in Neural Information Processing Systems , pages 3927—-3936, 2019. 8

    2019

  18. [25]

    RefQSR: Reference-based quantization for image super-resolution networks

    Hongjae Lee, Jun-Sang Yoo, and Seung-Won Jung. RefQSR: Reference-based quantization for image super-resolution networks. IEEE Transactions on Image Processing , 33: 2823–2834, 2024. 1, 3, 5

    2024

  19. [26]

    Auto- mated knowledge distillation via monte carlo tree search

    Lujun Li, Peijie Dong, Zimian Wei, and Ya Yang. Auto- mated knowledge distillation via monte carlo tree search. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17413–17424, 2023. 1

    2023

  20. [27]

    9 Q-diffusion: Quantizing diffusion models

    Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. 9 Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 17535–17545, 2023. 1, 2, 3, 5, 6, 8

    2023

  21. [28]

    BRECQ: Push- ing the limit of post-training quantization by block recon- struction

    Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. BRECQ: Push- ing the limit of post-training quantization by block recon- struction. In Proceedings of the International Conference on Learning Representations, 2021. 1, 5

    2021

  22. [29]

    Snap- Fusion: Text-to-image diffusion model on mobile devices within two seconds

    Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snap- Fusion: Text-to-image diffusion model on mobile devices within two seconds. In Proceedings of the Advances in Neu- ral Information Processing Systems , pages 20662–20678,

  23. [30]

    Microsoft COCO: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In Proceedings of the European Conference on Computer Vi- sion, pages 740–755, 2014. 3, 5, 6

    2014

  24. [31]

    Progressive neural architecture search

    Chenxi Liu, Barret Zoph, Maxim Neumann, Jonathon Shlens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan Yuille, Jonathan Huang, and Kevin Murphy. Progressive neural architecture search. In Proceedings of the European Conference on Com- puter Vision, pages 19–34, 2018. 1

    2018

  25. [32]

    Pseudo nu- merical methods for diffusion models on manifolds

    Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo nu- merical methods for diffusion models on manifolds. In Pro- ceedings of the International Conference on Learning Rep- resentations, 2022. 1

    2022

  26. [33]

    Instance-aware dynamic neural network quantization

    Zhenhua Liu, Yunhe Wang, Kai Han, Siwei Ma, and Wen Gao. Instance-aware dynamic neural network quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12434–12443, 2022. 1, 3

    2022

  27. [34]

    DPM-Solver: A fast ODE solver for dif- fusion probabilistic model sampling in around 10 steps

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver: A fast ODE solver for dif- fusion probabilistic model sampling in around 10 steps. In Proceedings of the Advances in Neural Information Process- ing Systems, pages 5775–5787, 2022. 1

    2022

  28. [35]

    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 Pat- tern Recognition, pages 14297–14306, 2023. 2

    2023

  29. [36]

    Improved denoising diffusion probabilistic models

    Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In Proceedings of the International Conference on Machine Learning , pages 8162–8171, 2021. 2

    2021

  30. [37]

    CUTLASS, 2025

    NVIDIA. CUTLASS, 2025. https://github.com/ NVIDIA/cutlass. 7

    2025

  31. [38]

    Softmax bias correction for quantized generative models

    Nilesh Prasad Pandey, Marios Fournarakis, Chirag Patel, and Markus Nagel. Softmax bias correction for quantized generative models. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision Workshop , pages 1453–1458, 2023. 2

    2023

  32. [39]

    Notes on regression and inheritance in the case of two parents

    Karl Pearson. Notes on regression and inheritance in the case of two parents. Proceedings of the Royal Society of London, 58:240–242, 1895. 8

  33. [40]

    SDXL: Improving latent diffusion mod- els for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion mod- els for high-resolution image synthesis. In Proceedings of the International Conference on Learning Representations ,

  34. [41]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. In Proceedings of the International Conference ...

    2021

  35. [42]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10674–10685, 2022. 1, 2, 3, 5, 6

    2022

  36. [43]

    DreamBooth: Fine tuning text-to-image diffusion models for subject-driven generation

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. 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, pages 22500...

    2023

  37. [44]

    Memory- efficient personalization using quantized diffusion model

    Hyogon Ryu, Seohyun Lim, and Hyunjung Shim. Memory- efficient personalization using quantized diffusion model. In Proceedings of the European Conference on Computer Vi- sion, 2024. 2

    2024

  38. [45]

    Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding. In Proceedings of the Ad- vanc...

    2022

  39. [46]

    Progressive distillation for fast sampling of diffusion models

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In Proceedings of the In- ternational Conference on Learning Representations , 2022. 2

    2022

  40. [47]

    Improved techniques for training gans

    Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved techniques for training gans. In Proceedings of the Ad- vances in Neural Information Processing Systems , pages 2234––2242, 2016. 5

    2016

  41. [48]

    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 Vi- sion and Pattern Recognition , pages 1972–1981, 2023. 1, 2

    1972

  42. [49]

    Conceptual captions: A cleaned, hypernymed, im- age alt-text dataset for automatic image captioning

    Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, im- age alt-text dataset for automatic image captioning. In Pro- ceedings of the 56th Annual Meeting of the Association for Computational Linguistics, pages 2556–2565, 2018. 5, 6

    2018

  43. [50]

    Temporal dynamic quantization for dif- fusion models

    Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, and Eunhyeok Park. Temporal dynamic quantization for dif- fusion models. In Proceedings of the Advances in Neural Information Processing Systems, pages 48686–48698, 2023. 1, 2, 3, 5 10

    2023

  44. [51]

    Denois- ing diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denois- ing diffusion implicit models. In Proceedings of the Inter- national Conference on Learning Representations, 2021. 1, 2

    2021

  45. [52]

    The proof and measurement of associa- tion between two things

    Charles Spearman. The proof and measurement of associa- tion between two things. The American Journal of Psychol- ogy, 15(1):72—-101, 1961. 8

    1961

  46. [53]

    CHIP: Channel independence- based pruning for compact neural networks

    Yang Sui, Miao Yin, Yi Xie, Huy Phan, Saman Aliari Zonouz, and Bo Yuan. CHIP: Channel independence- based pruning for compact neural networks. In Proceedings of the Advances in Neural Information Processing Systems , pages 24604–24616, 2021. 1

    2021

  47. [55]

    Post-training quan- tization with progressive calibration and activation relaxing for text-to-image diffusion models

    Siao Tang, Xin Wang, Hong Chen, Chaoyu Guan, Zewen Wu, Yansong Tang, and Wenwu Zhu. Post-training quan- tization with progressive calibration and activation relaxing for text-to-image diffusion models. In Proceedings of the European Conference on Computer Vision, 2024. 2, 3

    2024

  48. [56]

    CABM: Content-aware bit map- ping for single image super-resolution network with large input

    Senmao Tian, Ming Lu, Jiaming Liu, Yandong Guo, Yurong Chen, and Shunli Zhang. CABM: Content-aware bit map- ping for single image super-resolution network with large input. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 1756–1765,

  49. [57]

    Bayesian bits: Unifying quantization and pruning

    Mart Van Baalen, Christos Louizos, Markus Nagel, Rana Ali Amjad, Ying Wang, Tijmen Blankevoort, and Max Welling. Bayesian bits: Unifying quantization and pruning. In Pro- ceedings of the Advances in Neural Information Processing Systems, pages 5741–5752, 2020. 5

    2020

  50. [58]

    Ex- ploring clip for assessing the look and feel of images

    Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Ex- ploring clip for assessing the look and feel of images. InPro- ceedings of the AAAI Conference on Artificial Intelligence , pages 2555–2563, 2023. 8

    2023

  51. [59]

    HAQ: Hardware-aware automated quantization with mixed precision

    Kuan Wang, Zhijian Liu, Yujun Lin, Ji Lin, and Song Han. HAQ: Hardware-aware automated quantization with mixed precision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 8612– 8620, 2019. 1, 3

    2019

  52. [60]

    ImageRe- ward: Learning and evaluating human preferences for text- to-image generation

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. ImageRe- ward: Learning and evaluating human preferences for text- to-image generation. In Proceedings of the Advances in Neu- ral Information Processing Systems Workshop, 2023. 1

    2023

  53. [61]

    Online knowledge distillation via mutual contrastive learning for visual recog- nition

    Chuanguang Yang, Zhulin An, Helong Zhou, Fuzhen Zhuang, Yongjun Xu, and Qian Zhang. Online knowledge distillation via mutual contrastive learning for visual recog- nition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10212–10227, 2023. 1

    2023

  54. [62]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836–3847, 2023. 1

    2023

  55. [63]

    proj in” which leverages the denoised image latent passed into the cross-attention block, (2) “at2.to v

    Xuechen Zhang, Zheng Li, Samet Oymak, and Jiasi Chen. Text-to-3d generative ai on mobile devices: Measurements and optimizations. In Proceedings of the Workshop on Emerging Multimedia Systems, pages 8–14, 2023. 1 11 Text Embedding Knows How to Quantize Text-Guided Diffusion Mo...

    2023

Pith tools

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