Pith. sign in

REVIEW 3 major objections 5 minor 34 references

ExpertGen: Training-Free Expert Guidance for Controllable Text-to-Face Generation

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Consistency models turn off-the-shelf face experts into plug-and-play controllers for text-to-face generation.

desk verdict Useful combination of LCM clean intermediates and off-the-shelf face experts for training-free face control; the attribute/age/segmentation numbers are circular because the evaluator is the same frozen expert that supplies the guidance gradient. read the letter →

arxiv 2505.17256 v1 pith:YBHHU57I submitted 2025-05-22 cs.CV

Liang Shi , Yun Fu This is my paper · ORCID
classification cs.CV
keywords training-freeguidancetext-to-facegenerationlatentconsistencymodelexpertidentitypreservationfacialattributecontrolageestimationmulti-expert
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 claims that the reason training-free expert guidance fails in face generation is that the intermediate predictions a diffusion model feeds to face experts are unrealistic and out-of-distribution, so the experts' gradients are unreliable. It proposes fixing this by swapping the base sampler to a latent consistency model, whose self-consistency property makes intermediate predictions look like real faces even early in denoising. With that fix, plus a text-guided warmup phase and gradient clipping, off-the-shelf face analyzers can steer identity, attributes, age, and segmentation with no task-specific training, and several experts can be combined by summing gradients. If right, this turns any face analysis model into a plug-and-play controller and removes the training overhead of task-specific conditioning approaches for face generation.

What carries the argument

The load-bearing object is the intermediate prediction $\hat{x}_{0|t}$ produced by a latent consistency model during each denoising step, which the paper argues is realistic and in-distribution enough for face experts to evaluate. ExpertGen feeds that prediction into an off-the-shelf face expert, computes a task-specific loss (embedding similarity, classification, regression, or dense prediction), and back-propagates the gradient to the noisy latent $z_t$ using the update $\bar{\epsilon}_\theta(z_t,t) = \epsilon_\theta(z_t,t) - w\sqrt{1-\bar{\alpha}_t}\,\text{clip}(\nabla_{z_t}\mathcal{L}_{\text{exp}}(\hat{x}^{\text{LCM}}_{0|t}), -\tau, \tau)$. The two supporting mechanisms are text-guided warmup, which keeps the earliest still-blurry steps free of expert gradients by using a textual description of the target, and element-wise gradient clipping, which prevents unstable expert predictions from derailing the trajectory.

What would settle it

A direct test is to run ExpertGen with the same warmup, clipping, and gradient schedule but replace the latent consistency model with a standard latent diffusion model. If the standard-LDM variant matches the reported identity similarity of 0.594, attribute accuracy of 0.808, age error of 1.83 years, and segmentation accuracy of 0.768, then the claim that consistency-model intermediate predictions are the cause of the improvement is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that latent consistency models are the key enabler for training-free expert guidance in text-to-face generation. Unlike standard latent diffusion models, which produce blurry, ambiguous intermediate predictions that face experts misread, LCMs map every noisy latent to a realistic, in-distribution estimate at every step; face recognition, attribute, age, and segmentation networks can therefore provide accurate back-propagated gradients. ExpertGen keeps the standard training-free guidance update but computes the expert loss on the LCM intermediate prediction, delays expert gradients until after a text-guided warmup phase, and clips gradient elements. The result is single-expert control that beats text-only and LDM-based guidance on all four tasks—identity similarity of 0.594, attribute accuracy of 0.808, age error of 1.83 years, and segmentation pixel accuracy of 0.768—and multi-expert combination by gradient addition for face editing and age progression.

Load-bearing premise

The load-bearing premise is that improving the preview image at each step actually improves the final image, because the expert only ever sees the preview and never the final output.

Editorial extensions

If this is right

  • Any off-the-shelf face analysis model becomes a plug-and-play condition controller without retraining; adding a new condition only requires choosing a loss and a prompt.
  • Multiple experts can be combined by summing gradients, so tasks like identity-preserving age progression or attribute editing run in one framework with no task-specific modules.
  • Because the approach works on two latent diffusion backbones with a consistency adapter, it inherits the control of future diffusion backbones as long as a consistency adapter is available.
  • Attribute accuracy improves from 0.483 with text-only guidance to 0.808 on the smaller backbone and from 0.720 to 0.853 on the larger one, so experts add what text cannot express even in stronger models.
  • The method is demonstrated with 8 to 16 denoising steps, so fine-grained control comes at modest sampling cost compared with training-based conditioning modules.

Reading between the lines

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

  • Beyond the paper, the same mechanism should generalize to other domains that have off-the-shelf discriminative experts, such as medical image classifiers or scene attribute detectors, provided a consistency model for that domain exists.
  • A testable extension is to apply the same LCM-intermediate-prediction guidance to non-face attributes where experts are known to rely on spurious correlations, predicting that adversarial overfitting is reduced but not eliminated; the paper's own limitation section notes spurious correlations like 'blonde hair' biasing gender.
  • Because the authors report age error below the age estimator's own error on real images, a natural next experiment is to check whether guidance exploits estimator bias rather than true age; the paper acknowledges this ceiling.
  • The identity evaluation in the appendix uses two additional recognition networks to show invariance; an independent human-rater study would settle whether the control reads as intended to humans, not just to experts.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes ExpertGen, a training-free framework for controllable text-to-face generation. It replaces the standard latent diffusion model with a latent consistency model so that intermediate predictions fed to off-the-shelf face experts (ArcFace, FaRL, MiVOLO, Segface) are more realistic and in-distribution. The guidance signal is the gradient of an expert-defined loss with respect to the noisy latent, optionally after a text-guided warmup phase and with element-wise gradient clipping. The authors report quantitative results for identity, facial attribute, age, and segmentation-map guidance on SD-v1.5 and SDXL, including multi-expert combinations, and compare against no guidance, text-only guidance, and LDM-based UGD.

Significance. If the central claim is established, the paper would offer a practical and general alternative to task-specific training for face conditioning: any face-analysis expert could be plugged into generation without retraining, and multiple experts could be combined additively. The identity result is the strongest part of the paper, as it is evaluated against an independent face-recognition model and compared with several personalization baselines. The paper also provides useful ablations of warmup length, guidance weight, and clipping threshold, and the LCM-intermediate-prediction insight is well motivated by the FID and t-SNE observations. However, the quantitative evidence for attribute, age, and segmentation control is currently weakened by the fact that the guidance loss and the evaluation metric are the same function of the same frozen expert, and the invariance analysis only addresses identity. The central claim is defensible but needs additional independent evaluation before the reported numbers can be interpreted as semantic alignment rather than optimization of the guiding expert.

major comments (3)
  1. [Table 1 and Sec. 4.2] The quantitative evidence for attribute, age, and segmentation guidance is circular: for each of these tasks the guidance loss and the evaluation metric are identical functions of the same frozen expert (FaRL, MiVOLO, Segface). The reported gains (80.8% attribute accuracy, 1.83-year age MAE, 76.8% pixel accuracy) therefore partly measure how well the generated image satisfies the very classifier used to steer it, rather than the intended semantic condition. The paper itself acknowledges that the age error is below the average error of the age estimator on real images (Sec. 4.2), which indicates that the metric is saturating. The central claim that ExpertGen enables precise control over attributes, age, and segmentation needs at least one independent probe of the final image for each affected task: a different expert model of the same task, a human rating, or a held-out attribute/parsing model not used for guidance. Without such a probe, Tables 2 and 4 support the weaker claim that the generator can match the guidance expert's loss, not that the final images satisfy the intended conditions.
  2. [Sec. 4.4.2 and Table 9] The invariance-to-evaluation-models analysis covers only identity guidance. The alternative face-recognition networks (Buffalo-l, Antelope-v2) provide a meaningful cross-check for the identity task, where the guidance and evaluation networks are also both face-recognition models. However, no analogous experiment is reported for attributes, age, or segmentation, even though those are exactly the tasks where the guidance and evaluation networks coincide (Table 1). As a result, the statement in Sec. 4.4.2 that "different expert models respond positively to our generated images" is not established for the tasks most at risk of over-optimization. Please add cross-evaluation with a different attribute classifier, a different age estimator, and a different face parser (or human evaluation) for the corresponding guidance tasks.
  3. [Eq. (4) and Sec. 3.2] The method back-propagates a loss computed on the intermediate prediction x_0|t through the noisy latent z_t, and the paper assumes without direct evidence that optimizing this loss improves the final generated image rather than only the preview. Because the LCM trajectory is short (16 or 8 steps), the preview may be close to the final output, but the manuscript does not report a diagnostic comparing expert loss on intermediate predictions with expert loss on final decoded images. A simple experiment measuring the chosen expert's loss on final images across guidance strengths, or ablating the step at which guidance begins while measuring final-image alignment with an independent model, would substantially strengthen the transfer claim and address the preview-specific artifact concern.
minor comments (5)
  1. [Eq. (4) and Eq. (5)] The notation for the intermediate prediction is inconsistent: Eq. (4) uses \hat{x}^{LDM}_{0|t} while Eq. (5) uses \hat{x}^{LCM}_{0|t}; please use a single symbol for the LCM prediction.
  2. [Figure 1] The caption says "T-SNE" but the correct spelling is "t-SNE"; also the subfigure captions use inconsistent capitalization (e.g., "T-SNE Face Feature Distribution" vs. "FID along Diffusion Path").
  3. [Tables 5 and 8] The accuracy and quality values for overlapping settings appear inconsistent between Table 5 (main text) and Table 8 (appendix): for example, w=200, tau=0.0005 gives accuracy 0.841/quality 4.812 in Table 5 but 0.853/4.983 in Table 8. Please clarify whether these tables use different base models, step counts, or evaluation subsets, and state this explicitly.
  4. [Sec. 4.2 and Tables 2-4] No sample sizes, seeds, or confidence intervals are reported for the quantitative comparisons. Since several differences are small (e.g., segmentation accuracy 0.768 vs. 0.695), please report the number of generated images and the variance across seeds for the main tables.
  5. [Sec. 4.1] The text states "guidance scales are set at w = 200, and gradient clipping thresholds are set at tau = 5e - 4"; the spacing in "5e - 4" should be fixed to "5e-4" for consistency with Table 8.

Circularity Check

1 steps flagged · score 6.0 of 10

Attribute, age, and segmentation metrics are computed by the same experts that provide guidance, so those control claims are partly measured with the objective itself; only identity receives a cross-expert check.

  1. fitted input called prediction [Sec. 4.1-4.2, Eq. 4, Table 1]
    "We use task-specific expert models to evaluate the alignment between the target and the generated image."

    For the attribute, age, and segmentation rows of Table 1, the evaluation metric is the same expert output that defines the guidance loss: attribute uses FaRL's I(yhat=y) while guiding with FaRL's -y log p; age uses MiVOLO's |a-agt| while guiding with the same |a-agt|; segmentation uses Segface's pixel accuracy while guiding with Segface's dense cross-entropy. Eq. 4 back-propagates these very losses through every guided denoising step, so the reported numbers measure how well the final image satisfies the guiding expert, not whether an independent model or human verifies the condition. The paper concedes the age metric is unreliable (Sec. 4.2: 1.83 years below MiVOLO's error on real images), and the only cross-expert check (Table 9) is for identity, not for attributes, age, or segmentation.

full rationale

The identity-guidance results are largely self-contained: Table 9 evaluates ID similarity with ArcFace, Buffalo-l, and Antelope-v2, so the guidance network is not the sole judge, and FaceScore provides an independent quality signal. However, for attribute, age, and segmentation control, the evaluation network is identical to the guidance network (Table 1). Because Eq. 4 directly minimizes that same expert loss on the intermediate predictions, the reported accuracy, age error, and pixel accuracy are post-optimization scores of the guiding expert rather than independent measurements of the intended semantic condition. The paper itself flags the age number as below MiVOLO's real-image error, which limits its reliability. No analogous cross-evaluation is provided for attributes, age, or segmentation, so the central multi-task control claim rests in part on metrics that reduce to the guidance objective. There are no load-bearing self-citations; the circularity is metric-objective identity, and it is partial because the identity result, FaceScore, and relative comparisons to text-only and LDM guidance retain independent content.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The quantitative results depend on three hyperparameters tuned on the reported accuracy and quality metrics, plus the assumption that expert gradients transfer from intermediate predictions to final images. No new physical or conceptual entities are introduced.

free parameters (3)
  • guidance weight w = 200 (default)
    Selected by hyperparameter sweep in Appendix A.1 (Tables 5 and 8); the guidance update in Eq. 5 scales the expert gradient by w, and the reported headline results use this tuned value.
  • gradient clipping threshold tau = 5e-4 (0.0005)
    Selected from the sweep in Table 8 as the optimal trade-off between accuracy and FaceScore quality; clipping is a core component of Eq. 5 and affects all results.
  • warmup threshold T_thre = 800 (guidance begins at step 3 of an 8-step schedule)
    Chosen via the warmup-length sweep in Table 7; it determines which early denoising steps use text-only guidance and is critical to the method's performance.
assumptions (5)
  • standard math Latent diffusion and consistency model equations (Eqs. 1-3) and DDIM sampling (Sec. 3.1) are taken from prior work.
    The paper builds on [17,18,22,26,27] without re-deriving these foundations.
  • domain assumption LCM intermediate predictions are in-distribution enough for face experts to give reliable gradients.
    Core premise of the paper; supported only by Fig. 1 with 100 images and qualitative t-SNE/FID evidence, not by a formal guarantee or large-scale study.
  • domain assumption Gradients of a loss on intermediate prediction x_0|t steer the final generated image when applied to z_t (Eq. 4).
    Inherited from UGD/TFG; the authors do not prove that optimizing the intermediate prediction transfers to the final output.
  • domain assumption Text-guided warmup can bring early predictions close enough to the target condition for later expert guidance to work.
    Ablations in Table 4 show that delayed LCM guidance without text can degrade accuracy, so warmup is a critical empirical premise.
  • domain assumption Off-the-shelf face experts provide meaningful measures of condition alignment.
    The method depends on ArcFace, FaRL, MiVOLO, and Segface; the paper itself notes that age estimator error sets a lower bound on achievable accuracy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ExpertGen: Training-Free Expert Guidance for Controllable Text-to-Face Generation." pith.science (2026). https://pith.science/paper/YBHHU57I

@misc{pith2026250517256,
  author       = {Pith},
  title        = {Pith review of: ExpertGen: Training-Free Expert Guidance for Controllable Text-to-Face Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YBHHU57I}},
  note         = {Machine review of arXiv:2505.17256}
}
read the original abstract

Recent advances in diffusion models have significantly improved text-to-face generation, but achieving fine-grained control over facial features remains a challenge. Existing methods often require training additional modules to handle specific controls such as identity, attributes, or age, making them inflexible and resource-intensive. We propose ExpertGen, a training-free framework that leverages pre-trained expert models such as face recognition, facial attribute recognition, and age estimation networks to guide generation with fine control. Our approach uses a latent consistency model to ensure realistic and in-distribution predictions at each diffusion step, enabling accurate guidance signals to effectively steer the diffusion process. We show qualitatively and quantitatively that expert models can guide the generation process with high precision, and multiple experts can collaborate to enable simultaneous control over diverse facial aspects. By allowing direct integration of off-the-shelf expert models, our method transforms any such model into a plug-and-play component for controllable face generation.

Figures

Figures reproduced from arXiv: 2505.17256 by the authors.

Figure 1
Figure 1. Evaluating the quality of intermediate predictions. (a) FID of face images at different DDIM steps (earliest steps omitted due to large values). While both models improve, LCM achieves higher image quality much faster. (b) Face recognition features of LCM’s intermediate predictions converge within early steps. (c) Low-quality early-step predictions lack distinctive facial features, and therefore form a multi-color c… view at source ↗
Figure 2
Figure 2. Qualitative and quantitative results of facial attribute guidance. (a) We select five attributes that are challenging to generate using text conditions alone, and demonstrate how ExpertGen effectively morphs the image across eight DDIM time steps to generate correct attributes (zoom in for details). (b) Average probability of successful generation before and after ExpertGen across all 40 Celeb-A [16] facial attribut… view at source ↗
Figure 3
Figure 3. Visualizations of ID guidance, segmentation map guidance, and age guidance. Conditions of [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualizations of multi-expert guidance. Given an ID target on the left, we simultaneously [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Average probability of successful gener [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 15 canonical work pages

  1. [1]

    Universal guidance for diffusion models

    Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 843–852, 2023

    work page 2023

  2. [2]

    Vggface2: A dataset for recognising faces across pose and age

    Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018), pages 67–74. IEEE, 2018

    2018

  3. [3]

    Dreamidentity: enhanced editability for efficient face-identity preserved image generation

    Zhuowei Chen, Shancheng Fang, Wei Liu, Qian He, Mengqi Huang, and Zhendong Mao. Dreamidentity: enhanced editability for efficient face-identity preserved image generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 1281–1289, 2024

    2024

  4. [4]

    Arcface: Additive angular margin loss for deep face recognition

    Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019

    2019

  5. [5]

    Diffusion models beat GANs on image syn- thesis

    Prafulla Dhariwal and Alex Nichol. Diffusion models beat GANs on image syn- thesis. In Advances in Neural Information Processing Systems 34 (NeurIPS 2021) , pages 8780–8794, 2021. URL https://proceedings.neurips.cc/paper/2021/hash/ 49ad23d1ec9fa4bd8d77d02681df5cfa-Abstract.html

    work page 2021

  6. [6]

    Lcm-lookahead for encoder-based text-to-image personalization

    Rinon Gal, Or Lichter, Elad Richardson, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Lcm-lookahead for encoder-based text-to-image personalization. InEuropean Conference on Computer Vision, pages 322–340. Springer, 2024

    work page 2024

  7. [7]

    Insightface: 2d and 3d face analysis project

    Jia Guo, Jiankang Deng, and collaborators. Insightface: 2d and 3d face analysis project. https://github.com/deepinsight/insightface, 2023

    work page 2023

  8. [8]

    Face-adapter for pre-trained diffusion models with fine-grained id and attribute control

    Yue Han, Junwei Zhu, Keke He, Xu Chen, Yanhao Ge, Wei Li, Xiangtai Li, Jiangning Zhang, Chengjie Wang, and Yong Liu. Face-adapter for pre-trained diffusion models with fine-grained id and attribute control. In European Conference on Computer Vision, pages 20–36. Springer, 2024

    work page 2024

Show all 34 references
  1. [9]

    Manifold preserving guided diffusion

    Yutong He, Yifan Jiang, Yilun Xu, Shuang Li, Hang Zhao, Zhuowen Tu, and Yong Jae Lee. Manifold preserving guided diffusion. In Proceedings of the 37th Conference on Neural Information Processing Systems (NeurIPS 2023) , 2023. URL https://arxiv.org/abs/ 2311.16424

    2023 arXiv

  2. [10]

    Denoising diffusion probabilistic mod- els

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic mod- els. In Advances in Neural Information Processing Systems 33 (NeurIPS 2020) , pages 6840–6851, 2020. URL https://proceedings.neurips.cc/paper/2020/file/ 4c5bcfec8584af0d967f1ab10179ca4b-Paper.pdf

    2020

  3. [11]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1 (2):3, 2022

    2022

  4. [12]

    Enhancing diffusion-based image synthesis with robust classifier guidance

    Bahjat Kawar, Roy Ganz, and Michael Elad. Enhancing diffusion-based image synthesis with robust classifier guidance. arXiv preprint arXiv:2208.08664, 2022

    2022 arXiv

  5. [13]

    Jain, and Xiaoming Liu

    Minchul Kim, Anil K. Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  6. [14]

    Mivolo: Multi-input transformer for age and gender estimation

    Maksim Kuprashevich and Irina Tolstykh. Mivolo: Multi-input transformer for age and gender estimation. In International Conference on Analysis of Images, Social Networks and Texts , pages 212–226. Springer, 2023

    2023

  7. [15]

    Facescore: Benchmarking and enhancing face quality in human generation

    Zhenyi Liao, Qingsong Xie, Chen Chen, Hannan Lu, and Zhijie Deng. Facescore: Benchmarking and enhancing face quality in human generation. arXiv preprint arXiv:2406.17100, 2024

    2024 arXiv

  8. [16]

    Deep learning face attributes in the wild

    Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), December 2015

    2015

  9. [17]

    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. 10

    2023 arXiv

  10. [18]

    Lcm-lora: A universal stable-diffusion acceleration module

    Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023

    2023 arXiv

  11. [19]

    Segface: Face segmentation of long-tail classes

    Kartik Narayan, Vibashan Vs, and Vishal M Patel. Segface: Face segmentation of long-tail classes. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 6182–6190, 2025

    2025

  12. [20]

    Precisecontrol: Enhancing text-to-image diffusion models with fine-grained attribute control

    Rishubh Parihar, VS Sachidanand, Sabariswaran Mani, Tejan Karmali, and R Venkatesh Babu. Precisecontrol: Enhancing text-to-image diffusion models with fine-grained attribute control. In European Conference on Computer Vision, pages 469–487. Springer, 2024

    2024

  13. [21]

    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

  14. [22]

    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 (CVPR), pages 10684–10695, 2022. URL https://ar...

    2022 arXiv

  15. [23]

    Stable diffusion, 2022

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. Stable diffusion, 2022. URL https://github.com/CompVis/stable-diffusion

    2022

  16. [24]

    Face2diffusion for fast and editable face personaliza- tion

    Kaede Shiohara and Toshihiko Yamasaki. Face2diffusion for fast and editable face personaliza- tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6850–6859, 2024

    2024

  17. [25]

    Robustness and generalizability of deepfake detection: A study with diffusion models

    Haixu Song, Shiyu Huang, Yinpeng Dong, and Wei-Wei Tu. Robustness and generalizability of deepfake detection: A study with diffusion models. arXiv preprint arXiv:2309.02218, 2023

    2023 arXiv

  18. [26]

    Denoising diffusion implicit models

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

    2010 arXiv

  19. [27]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. 2023

    2023

  20. [28]

    Rectifid: Personalizing rectified flow with anchored classifier guidance

    Zhicheng Sun, Zhenhao Yang, Yang Jin, Haozhe Chi, Kun Xu, Liwei Chen, Hao Jiang, Yang Song, Kun Gai, and Yadong Mu. Rectifid: Personalizing rectified flow with anchored classifier guidance. Advances in Neural Information Processing Systems, 37:96993–97026, 2024

    2024

  21. [29]

    Visualizing data using t-sne

    Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008

    2008

  22. [30]

    TFG: Unified training-free guidance for diffusion models

    Haotian Ye, Haowei Lin, Jiaqi Han, Minkai Xu, Sheng Liu, Yitao Liang, Jianzhu Ma, James Zou, and Stefano Ermon. TFG: Unified training-free guidance for diffusion models. In Advances in Neural Information Processing Systems 37 (NeurIPS 2024), 2024. URL https: //openreview.net/f...

    2024

  23. [31]

    Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models

    Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023

    2023 arXiv

  24. [32]

    Freedom: Training- free energy-guided conditional diffusion model

    Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training- free energy-guided conditional diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23174–23184, 2023

    2023

  25. [33]

    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

    2023

  26. [34]

    General facial representation learning in a visual- linguistic manner

    Yinglin Zheng, Hao Yang, Ting Zhang, Jianmin Bao, Dongdong Chen, Yangyu Huang, Lu Yuan, Dong Chen, Ming Zeng, and Fang Wen. General facial representation learning in a visual- linguistic manner. In Proceedings of the IEEE/CVF conference on computer vision and pattern recogniti...

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.