Pith. sign in

REVIEW 4 major objections 5 minor 29 references

StyleAutoEncoder for manipulating image attributes using pre-trained StyleGAN

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read StyleAE is a lightweight AutoEncoder plugin for a pre-trained StyleGAN that edits image attributes in the style space with accuracy comparable to normalizing-flow methods while preserving unrelated image features.

desk verdict The method is plausible and cheap, but the headline claim is not yet supported because the evaluation may be measuring its own steering classifier. read the letter →

arxiv 2412.20164 v1 pith:5QOY5FAU submitted 2024-12-28 cs.CV cs.AI

Andrzej Bedychaj , Jacek Tabor , Marek Śmieja This is my paper · ORCID
classification cs.CVcs.AI
keywords StyleGANlatentspacemanipulationautoencoderpluginattributeeditingnormalizingflowsdisentangledrepresentationfaceimagegeneration
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

StyleAutoEncoder (StyleAE) is a plugin module for a pre-trained StyleGAN that edits a requested image attribute by re-encoding the StyleGAN style vector $w$ into a target space in which one coordinate corresponds to each labelled attribute, then decoding modified coordinates back to a style vector. The paper claims that this simple autoencoder approach matches the attribute-modification accuracy of normalizing-flow plugins StyleFlow and PluGeN while being markedly faster to train and run and less invasive to unrelated image features. The claim matters because flow-based disentanglement methods are computationally heavy and sensitive to hyperparameters, and training a full generative model from scratch is expensive; a lightweight plugin offers attribute control on limited compute. The paper supports the claim with quantitative experiments on human faces (FFHQ) and a qualitative animal-face study (AFHQv2).

What carries the argument

The central object is the StyleAE autoencoder defined on StyleGAN's style space $W$: an encoder $E:W\to(C,S)$ and a decoder $D:(C,S)\to W$, with no dimensionality reduction. The target space separates labelled attributes into individual coordinates $c_k$ while unlabelled variation is carried by $s_m$. The key mechanism is the two-part loss: an image loss $\|x-G(D(E(w)))\|_2$ keeps edits anchored to the original image, and an attribute loss aligns each $c_k$ with its label, with the asymmetric positive-label term $\max(0,1-c_k)$ allowing different positive styles rather than a single value. It is this loss, rather than an invertibility constraint, that produces the edit directions.

What would settle it

Run the FFHQ attribute-manipulation comparison with two disjoint classifiers: one to set the minimal-modification stopping point and another, independently trained classifier to report the accuracy in Table 2; if StyleAE's accuracy drops below StyleFlow or PluGeN on several attributes, the at-least-comparable claim fails.

Watch

Extended reading notes

Core claim

The central claim is that an AutoEncoder trained on StyleGAN's style vectors can produce a latent representation in which attributes are separable enough for editing, without any invertibility constraint. StyleAE's encoder maps a pre-trained StyleGAN style vector $w$ to coordinates $(c_1,...,c_K,s_1,...,s_M)$; the decoder maps back to $w$. Training minimizes image reconstruction loss plus an attribute alignment loss between the $c_k$ coordinates and the labels $y_k$, using the margin term $\max(0,1-c_k)$ for positive binary labels. At test time, editing shifts the relevant coordinate and decodes; the paper reports accuracy comparable to StyleFlow and PluGeN on FFHQ attributes, lower perceptual MSE, better structural preservation on many attributes, and training and inference times that are a fraction of the flow-based methods.

Load-bearing premise

The comparison assumes that the classifier that decides when a minimal attribute modification is sufficient is not the same classifier whose accuracy is reported in Table 2, since otherwise the reported accuracy partly measures how well the search satisfies that classifier rather than an independent attribute change.

Editorial extensions

If this is right

  • Attribute manipulation can be added to a pre-trained StyleGAN with only a few fully connected layers, so controllable generation becomes feasible on a single consumer GPU.
  • StyleAE preserves more of the input image's identity-related structure than flow plugins, as measured by perceptual MSE, indicating more localized edits.
  • Because the encoder-decoder does not need to be invertible, the architecture can be freely changed or deepened without altering the plugin's design.
  • The paper argues the approach applies to arbitrary generative models, so the same plugin recipe can be reused with other generators, not only StyleGAN.
  • The asymmetric binary loss stores different realizations of a positive attribute, which yields diverse outputs for edits such as beard or glasses.

Reading between the lines

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

  • Editorial inference: the same objective could be transplanted to latent spaces of diffusion or transformer generators, where invertible flow plugins are unavailable, provided a latent-to-image reconstruction path exists.
  • Editorial inference: because the loss is modular, a perceptual or identity-preservation term could be added directly in target space, a change flow-based methods cannot make without breaking invertibility.
  • Editorial inference: the AFHQv2 setup suggests a practical recipe for new domains: project real images into the generator's latent space, label them with a lightweight classifier, and train the plugin; a natural test is measuring editing accuracy on human-annotated labels.
  • Editorial inference: a two-classifier evaluation, one to steer the minimal edit and one to score the result, would give a cleaner comparison and is directly testable with the authors' code.
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 manuscript proposes StyleAutoEncoder (StyleAE), a lightweight AutoEncoder module that learns a mapping between StyleGAN's style space and a target space in which individual coordinates correspond to labelled attributes. The method trains only an encoder/decoder pair on image-attribute pairs, then manipulates attributes by shifting the corresponding coordinate and decoding back to style space. The authors compare StyleAE against the flow-based baselines StyleFlow and PluGeN on FFHQ and AFHQv2, reporting attribute-modification accuracy, reconstruction/preservation metrics, and training/inference times. The central claim is that StyleAE is at least as effective as flow-based state-of-the-art methods at manipulating attributes while being simpler, faster, and more architecture-flexible.

Significance. If the central claim were established, the contribution would be practically significant: a cheap, plug-in module that gives attribute control over a pre-trained StyleGAN would lower the computational barrier for controlled generation, and the AutoEncoder formulation would indeed allow more flexible losses and architectures than invertible-flow alternatives. The paper also has concrete engineering strengths: it builds on a pre-trained StyleGAN without retraining it, uses a simple three-layer encoder/decoder, trains in roughly 100 epochs, and reports substantially lower training and inference times than the flow baselines. The qualitative examples in Figs. 3 and 4 are visually plausible, and the AFHQv2 experiment extends the idea to an animal domain with projected latent codes. However, the comparative evaluation has a load-bearing circularity and several text claims are contradicted by the paper's own tables, so the significance is conditional on fixing the evaluation protocol.

major comments (4)
  1. [§3.6, Tables 1, 2 and 4] The minimal-modification procedure uses an attribute classifier to stop the traversal: 'we gradually modify the attribute until the classifier recognizes the attribute of the generated image with sufficient confidence.' Table 2 then reports 'the classifier's predictive accuracy' for the same attributes, and Tables 1 and 4 are computed on these classifier-stopped images. Unless the classifier used to choose the stopping point is provably a different network from the one used to score the result, the reported accuracy partly measures the search's ability to satisfy the scoring classifier, not an independent verification that the semantic attribute changed. Section 3.4 promises 'an independent multi-label face attribute classifier,' but Section 3.6 does not state that the stopping classifier and the evaluation classifier are distinct. This circularity is load-bearing because the central comparative claim rests on Table 2, and it propagates to the preservation metrics through the selection of minimally modified images. Please clarify whether the classifiers are distinct, and if they are not, rerun the evaluation with a held-out classifier that is not used in any part of the modification search.
  2. [§3.6, Table 1 and surrounding text] The text states that 'StyleAE obtains significantly lower MSE than state-of-the-art models' for perceptual MSE, but Table 1 shows the opposite for several attributes: for 'man' StyleAE has 73.241 vs 56.955 and 53.999 for PluGeN and StyleFlow; for 'woman' 36.650 vs 28.300 and 23.507; for 'no smile' 47.496 vs 36.251 and 36.471. At minimum the sentence 'significantly lower MSE' is false on these rows and should be corrected to an attribute-by-attribute description that acknowledges where StyleAE is worse. This matters because the abstract's 'at least as effective' claim is supported partly by preservation quality, and an overclaim in the main text needs fixing.
  3. [§3.5, §3.6, Table 3] The baseline comparison is not apples-to-apples. Section 3.5 says 'We use publicly available checkpoints for evaluation, avoiding retraining PluGeN or StyleFlow ourselves,' while Section 3.6 says 'All considered methods were trained on 10 000 images generated by StyleGAN.' If the flow baselines were trained on their original datasets and not on the same 10,000 generated images with the same attribute labels, then the accuracy and preservation comparisons in Tables 1, 2, and 4 may reflect training-data and label differences rather than method differences. The training-time comparison in Table 3 is also ambiguous: StyleAE is reported as 100 epochs of training, while the table's 'Training' row is described as 'Average time of 1 training epoch,' so the total training cost of the baselines is not stated. Please state explicitly which checkpoints were used, whether they were retrained under the same protocol, and report total training time to convergence for all methods.
  4. [Tables 2, 3 and 4] The tables report single numbers with no error bars, no confidence intervals, and no description of the number of test images or repeated trials. Given that several differences are small (e.g., 'no glasses' accuracy is 0.74 for StyleAE vs 0.78 for PluGeN, and 'no smile' is 0.99 vs 1.0), the claim that StyleAE is 'at least as effective' is statistically unsupported. Please add means and standard deviations across multiple seeds or bootstrap confidence intervals, and specify the test-set size used for each table.
minor comments (5)
  1. [§3.6, §3.4] The evaluation classifier is described as a ResNet18 trained for 8 target classes, but no training/test split, dataset source, or code link is given. Please provide the classifier's training details and a public link to enable reproducibility.
  2. [§3.7, Table 5 caption] The caption contains a typo: 'Finner traversing' should be 'Finer traversing'. In addition, the caption's admission that 'the style transfer is not ideal' should be reconciled with the text's claim that the results 'attest to the robustness and effectiveness' of the method; the qualitative evidence is weaker than the prose suggests.
  3. [§3.3 and §3.6] The paper says in §3.3 that the human facial feature dataset was 'categorized externally using the Microsoft Face API,' while §3.6 says a new classifier was developed because of Microsoft API licensing changes. It would help to state explicitly that Microsoft Face API was used only for generating training labels and that the new classifier is used only for evaluation, if that is the case.
  4. [Fig. 1 and Fig. 4 captions] The captions are visually noisy (e.g., 'Input Old Beard Gender Input No Glasses Y oung Smile' in Fig. 1) and would benefit from spacing and punctuation. Also, Fig. 4's caption says 'The generated images by all models exhibit successful changes,' which is a stronger claim than the quantitative tables support; please align the caption with the reported numbers.
  5. [§3.1, Eq. (1)] The notation z is used both for the standard Gaussian latent vector of StyleGAN and for the AutoEncoder latent representation in Eq. (1). Consider using a different symbol for one of them to avoid confusion.

Circularity Check

1 steps flagged · score 6.0 of 10

Table 2 accuracy is the same classifier used to stop the minimal-modification traversal, so the central comparative claim partly reduces to the stopping rule.

  1. self definitional [Section 3.6, Table 2 caption]
    "Since every method can use different scales to represent the intensity of attributes being modelled, we employed an attribute classifier to apply a minimal modification to obtain the requested value of the attribute. In other words, we gradually modify the attribute until the classifier recognizes the attribute of the generated image with sufficient confidence. If we cannot obtain the requested modification, the classifier returns failure."

    The success metric in Table 2 is the predictive accuracy of the same attribute classifier that defines the stopping rule of the modification. Because the traversal continues until that classifier recognizes the attribute, the reported accuracy is essentially the success rate of the stopping rule against its own oracle: a sample counts as successfully modified only when the classifier says so. Section 3.4's 'independent' claim refers to the classifier's training data, not to independence from the evaluation protocol, and no separate scoring classifier is introduced. Hence Table 2 does not measure a semantically independent attribute change; it measures the method's ability to satisfy the classifier, which was built into the stopping procedure.

full rationale

The paper's core derivation--training StyleAE with an attribute loss (Eq. 3) so that latent coordinates align with labeled attributes--is not circular: it is a standard supervised objective, and the resulting edit directions are genuinely learned from labels rather than defined by the evaluation. The main circularity is evaluative and located in the accuracy protocol. Section 3.6 uses one attribute classifier to decide when a 'minimal modification' is sufficient and then reports that same classifier's recognition as the accuracy in Table 2. This makes the accuracy numbers, and the headline comparative claim built on them, partly a measure of the stopping rule's success against its own oracle rather than an independent verification of semantic attribute editing. The preservation metrics (MSE, PSNR, SSIM, perceptual MSE) are computed on images whose stopping point is nevertheless chosen by that classifier, so the circularity propagates to those comparisons as well, though less directly. No other circular steps were found: the baselines use publicly available checkpoints, self-citations to PluGeN work are not load-bearing for the method, and no uniqueness theorem or ansatz is smuggled in via citation.

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

The central claim is empirical: StyleAE is a trained mapping with several hand-set hyperparameters. Its feasibility rests on the capacity of a small MLP to disentangle W-space and on the reliability of attribute labels and projection methods. No new physical entities are introduced.

free parameters (3)
  • attribute loss weight (max) = 0.3, annealed from 0 over first 30 epochs
    Balances attribute alignment against image reconstruction; chosen by hand with no reported cross-validation.
  • classifier confidence threshold = not specified
    The 'minimal modification' search stops when the attribute classifier reports sufficient confidence; the exact threshold is unstated and directly controls accuracy and preservation metrics in Tabs. 1, 2, 4.
  • target space dimensionality = 512 (no compression)
    The paper keeps the target latent space at the same 512 dimensions as W for fair comparison with flows; this is a design choice, not a learned value.
assumptions (4)
  • standard math Backpropagation and gradient descent can minimize the composite image and attribute loss
    Standard assumption for training neural networks; no formal convergence guarantee needed for empirical claims.
  • domain assumption A three-layer MLP autoencoder on StyleGAN's W-space can learn an approximately invertible mapping with axis-aligned attribute coordinates
    Section 3.2 defines the encoder and decoder as 512-neuron MLPs; the central feasibility claim depends on this capacity assumption.
  • domain assumption The Microsoft Face API labels used for training and the authors' ResNet18 classifier provide accurate proxies for the eight target attributes
    Section 3.6 trains on Face API labels and uses the ResNet18 classifier for modification and evaluation; label noise would propagate into all metrics.
  • domain assumption The image projection method of Abdal et al. [1] yields style vectors that faithfully represent AFHQv2 images
    Section 3.7 trains StyleAE on projected real animal images; if the projections are poor, the qualitative style transfer results on AFHQv2 would not transfer to real images.

how reviews work

0 comments
Cite this review

Pith. "Pith review of StyleAutoEncoder for manipulating image attributes using pre-trained StyleGAN." pith.science (2026). https://pith.science/paper/5QOY5FAU

@misc{pith2026241220164,
  author       = {Pith},
  title        = {Pith review of: StyleAutoEncoder for manipulating image attributes using pre-trained StyleGAN},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5QOY5FAU}},
  note         = {Machine review of arXiv:2412.20164}
}
read the original abstract

Deep conditional generative models are excellent tools for creating high-quality images and editing their attributes. However, training modern generative models from scratch is very expensive and requires large computational resources. In this paper, we introduce StyleAutoEncoder (StyleAE), a lightweight AutoEncoder module, which works as a plugin for pre-trained generative models and allows for manipulating the requested attributes of images. The proposed method offers a cost-effective solution for training deep generative models with limited computational resources, making it a promising technique for a wide range of applications. We evaluate StyleAutoEncoder by combining it with StyleGAN, which is currently one of the top generative models. Our experiments demonstrate that StyleAutoEncoder is at least as effective in manipulating image attributes as the state-of-the-art algorithms based on invertible normalizing flows. However, it is simpler, faster, and gives more freedom in designing neural

Figures

Figures reproduced from arXiv: 2412.20164 by the authors.

Figure 1
Figure 1. Single-attribute manipulation with StyleAE in StyleGAN’s latent space. 1 Introduction Generative models, such as generative adversarial networks (GAN) [10], varia￾tional AutoEncoders (VAE) [20], diffusion models [23], and flow-based genera￾tive models [8], have gained popularity due to their ability to create high-quality images, videos, and texts. These models are trained using supervised or unsu￾pervised learning … view at source ↗
Figure 2
Figure 2. Architecture design of StyleAE. StyleAE maps the style code w of the pre-trained StyleGAN into a target space, where labelled attributes are modelled by individual coordinates. AutoEncoder: architecture comprises an encoder function, mapping input data to a compressed representation in the latent space, and a decoder function, mapping this representation back to the original input space. This is represented as: z = … view at source ↗
Figure 3
Figure 3. Examples of attributes modification for all of the tested models on [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Attribute modification on a sample image generated from StyleGAN. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 19 canonical work pages

  1. [1]

    Image2StyleGAN: How to Embed Images Into the StyleGAN Latent Space?

    Abdal, R., Qin, Y., Wonka, P.: Image2stylegan: How to embed images into the stylegan latent space? vol. abs/1904.03189 (2019) 6, 10 12 A. Bedychaj et al

    work page Pith review arXiv 2019

  2. [2]

    CoRR abs/2112.05219 (2021) 5

    Abdal, R., Zhu, P., Femiani, J., Mitra, N.J., Wonka, P.: Clip2stylegan: Unsuper- vised extraction of stylegan edit directions. CoRR abs/2112.05219 (2021) 5

    arXiv 2021

  3. [3]

    StyleFlow: Attribute-conditioned Exploration of StyleGAN-Generated Images using Conditional Continuous Normalizing Flows

    Abdal, R., Zhu, P., Mitra, N.J., Wonka, P.: Styleflow: Attribute-conditioned ex- ploration of stylegan-generated images using conditional continuous normalizing flows. CoRR abs/2008.02401 (2020) 2, 3, 6, 8

    work page Pith review arXiv 2020

  4. [4]

    Cha, J., Thiyagalingam, J.: Disentangling autoencoders (dae) (2022) 7

    work page 2022

  5. [5]

    Chen, X., Duan, Y., Houthooft, R., Schulman, J., Sutskever, I., Abbeel, P.: Info- gan: Interpretable representation learning by information maximizing generative adversarial nets (2016) 2

    work page 2016

  6. [6]

    CoRR abs/1912.01865 (2019) 7

    Choi, Y., Uh, Y., Yoo, J., Ha, J.: Stargan v2: Diverse image synthesis for multiple domains. CoRR abs/1912.01865 (2019) 7

    arXiv 2019

  7. [7]

    Deng, J., Guo, J., Zafeiriou, S.: Arcface: Additive angular margin loss for deep face recognition. vol. abs/1801.07698 (2018) 7

    arXiv 2018

  8. [8]

    Dinh, L., Sohl-Dickstein, J., Bengio, S.: Density estimation using real NVP (2016) 1

    work page 2016

Show all 29 references
  1. [9]

    CoRR abs/2103.15814 (2021) 1, 3

    Gao, Y., Wei, F., Bao, J., Gu, S., Chen, D., Wen, F., Lian, Z.: High-fidelity and arbitrary face editing. CoRR abs/2103.15814 (2021) 1, 3

    2021 arXiv

  2. [10]

    Goodfellow, I.J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks (2014) 1

    2014

  3. [11]

    CoRR abs/2004.02546 (2020) 2, 3

    H¨ ark¨ onen, E., Hertzmann, A., Lehtinen, J., Paris, S.: Ganspace: Discovering inter- pretable GAN controls. CoRR abs/2004.02546 (2020) 2, 3

    2020 arXiv

  4. [12]

    CoRR abs/1512.03385 (2015) 8

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. CoRR abs/1512.03385 (2015) 8

    2015 arXiv

  5. [13]

    IEEE Transactions on Image Processing 28(11), 5464–5478 (Nov 2019) 1, 3

    He, Z., Zuo, W., Kan, M., Shan, S., Chen, X.: Attgan: Facial attribute editing by only changing what you want. IEEE Transactions on Image Processing 28(11), 5464–5478 (Nov 2019) 1, 3

    2019

  6. [14]

    CoRR abs/1902.00275 (2019) 2

    Ho, J., Chen, X., Srinivas, A., Duan, Y., Abbeel, P.: Flow++: Improving flow- based generative models with variational dequantization and architecture design. CoRR abs/1902.00275 (2019) 2

    2019 arXiv

  7. [15]

    In: Proc

    Karras, T., Aittala, M., Laine, S., H¨ ark¨ onen, E., Hellsten, J., Lehtinen, J., Aila, T.: Alias-free generative adversarial networks. In: Proc. NeurIPS (2021) 1, 3

    2021

  8. [16]

    CoRR abs/1812.04948 (2018) 1, 2, 3, 7

    Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. CoRR abs/1812.04948 (2018) 1, 2, 3, 7

    2018 arXiv

  9. [17]

    CoRR abs/1912.04958 (2019) 1, 3

    Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. CoRR abs/1912.04958 (2019) 1, 3

    2019 arXiv

  10. [18]

    Kingma, D.P., Dhariwal, P.: Glow: Generative flow with invertible 1x1 convolutions (2018) 2

    2018

  11. [19]

    CoRR abs/1406.5298 (2014) 3

    Kingma, D.P., Rezende, D.J., Mohamed, S., Welling, M.: Semi-supervised learning with deep generative models. CoRR abs/1406.5298 (2014) 3

    2014 arXiv

  12. [20]

    arXiv preprint arXiv:1312.6114 (2013) 1

    Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013) 1

    2013 arXiv

  13. [21]

    CoRR abs/1706.00409 (2017) 3

    Lample, G., Zeghidour, N., Usunier, N., Bordes, A., Denoyer, L., Ran- zato, M.: Fader networks: Manipulating images by sliding attributes. CoRR abs/1706.00409 (2017) 3

    2017 arXiv

  14. [22]

    CoRR abs/1904.01782 (2019) 1, 3

    Liu, R., Liu, Y., Gong, X., Wang, X., Li, H.: Conditional adversarial generative flow for controllable image synthesis. CoRR abs/1904.01782 (2019) 1, 3

    2019 arXiv

  15. [23]

    In: CVPR (2022) 1

    Preechakul, K., Chatthee, N., Wizadwongsa, S., Suwajanakorn, S.: Diffusion au- toencoders: Toward a meaningful and decodable representation. In: CVPR (2022) 1

    2022

  16. [24]

    CoRR abs/2005.09635 (2020) 1, 2, 3 StyleAutoEncoder 13

    Shen, Y., Yang, C., Tang, X., Zhou, B.: Interfacegan: Interpreting the disentangled face representation learned by gans. CoRR abs/2005.09635 (2020) 1, 2, 3 StyleAutoEncoder 13

    2020 arXiv

  17. [25]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Suwa la, A., W´ ojcik, B., Proszewska, M., Tabor, J., Spurek, P., ´Smieja, M.: Face identity-aware disentanglement in stylegan. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 5222–5231 (2024) 2, 3

    2024

  18. [26]

    ACM Trans

    Tewari, A., Elgharib, M., R, M.B., Bernard, F., Seidel, H.P., P´ erez, P., Zollh¨ ofer, M., Theobalt, C.: Pie: Portrait image embedding for semantic control. ACM Trans. Graph. 39(6) (2020) 1, 3

    2020

  19. [27]

    Scientific Reports 13 (03 2023) 2

    Vidal, A., Wu Fung, S., Tenorio, L., Osher, S., Nurbekyan, L.: Taming hyperpa- rameter tuning in continuous normalizing flows using the jko scheme. Scientific Reports 13 (03 2023) 2

    2023

  20. [28]

    CoRR abs/2011.14107 (2020) 1, 3

    Wang, H., Yu, N., Fritz, M.: Hijack-gan: Unintended-use of pretrained, black-box gans. CoRR abs/2011.14107 (2020) 1, 3

    2020 arXiv

  21. [29]

    In: AAAI 2022 (2022) 2, 3, 6, 8

    Wo lczyk, M., Proszewska, M., Maziarka, L., Zieba, M., Wielopolski, P., Kurczab, R., ´Smieja, M.: Plugen : multi-label conditional generation from pre-trained mod- els. In: AAAI 2022 (2022) 2, 3, 6, 8

    2022

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.