Pith. sign in

REVIEW 4 major objections 5 minor 41 references

Towards Fair and Robust Face Parsing for Generative AI: A Multi-Objective Approach

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read The paper claims that face parsers trained with a homotopy-weighted balance of accuracy, fairness, and robustness produce segmentation maps that make downstream GAN-based face synthesis more photorealistic and demographically consistent…

desk verdict The paper's central robustness claim is unsupported because the training loss perturbs logits while the evaluation perturbs inputs, and the mIoU losses are not differentiable as written; Table III also looks like placeholder data. read the letter →

arxiv 2502.04391 v1 pith:LONPMS3C submitted 2025-02-06 cs.CV cs.AI

classification cs.CVcs.AI
keywords faceparsingmulti-objectivelearningfairnessrobustnesshomotopyoptimizationGAN-basedsynthesisdiffusionmodelssemanticsegmentation
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

Face parsing—labeling every part of a face (hair, eyes, nose, mouth) in an image—is usually trained for accuracy alone, and that accuracy can hide uneven performance across demographic groups and under noise, blur, or occlusion. The authors propose a multi-objective training scheme that optimizes accuracy, fairness, and robustness at once, using a homotopy-based loss that shifts emphasis from accuracy early in training to the other objectives later. Their central claim is that the resulting segmentation maps improve downstream generative models: the same Pix2PixHD GAN produces lower FID and LPIPS scores when conditioned on multi-objective maps than on single-objective maps, and preliminary ControlNet experiments move in the same direction. The paper thus argues that fairness and robustness can be introduced at the segmentation stage, before biases reach the generator.

What carries the argument

The central mechanism is a homotopy-based multi-objective loss, $L_{\text{total}} = \alpha L_{\text{acc}} + \beta L_{\text{rob}} + \gamma L_{\text{fair}}$ with $\alpha + \beta + \gamma = 1$, whose weights change by epoch under linear, sigmoid, or piecewise schedules. Here $L_{\text{acc}}$ is Dice loss against the ground-truth mask, $L_{\text{fair}}$ is the variance of per-group mIoU across demographic attributes, and $L_{\text{rob}}$ is the negative mIoU computed on the network's output logits after adding random noise. The homotopy schedule matters because it lets training start accuracy-first and gradually deform toward fairness and robustness, which the authors argue prevents early multi-objective interference. These losses are engineered inside a U-Net parser, and the segmentation maps it emits are then consumed by Pix2PixHD and ControlNet.

What would settle it

Train two parsers with identical multi-objective schedules except that one uses the paper's logit-noise robustness loss and the other is trained directly on input perturbations; if the logit-noise parser shows no better mIoU under the paper's input-perturbation test than an accuracy-only baseline, the claimed robustness mechanism is not what produces the reported gains.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that adding fairness and robustness objectives to face parsing does not force a trade-off against accuracy. A U-Net trained with the multi-objective loss matches or slightly improves mean IoU relative to a single-objective baseline (74.21% vs. 73.87% for the linear schedule) while reducing variance across demographic groups and retaining more coherent segmentation under Gaussian noise, blur, and occlusion. When these segmentation maps are fed to the same Pix2PixHD generator, FID improves from 117.93 for the baseline parser to 98.87 for the piecewise multi-objective parser, and LPIPS drops from 0.4419 to 0.4222. The authors interpret the combination as evidence that fairness- and robustness-aware segmentation is a viable way to obtain higher-quality and more equitable generative face synthesis.

Load-bearing premise

The argument assumes that making the parser's output logits resistant to added random noise is enough to make the parser robust to input-level noise, blur, and occlusion, because the training procedure never applies those input perturbations.

Editorial extensions

If this is right

  • Fairness and robustness can be added to face parsing without sacrificing accuracy: all multi-objective schedules stay within roughly one percentage point of the single-objective mIoU baseline.
  • The same GAN generator produces better images when its conditioning maps come from a multi-objective parser, indicating that upstream segmentation quality is a lever on generative quality.
  • Reducing demographic variance at the parser level is a way to reduce bias in generated faces without retraining or altering the generator architecture.
  • The ControlNet results, though limited to one epoch, suggest the benefit extends to diffusion-based conditioning, not just GANs.

Reading between the lines

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

  • A direct test of the robustness claim would replace the logit-noise robustness loss with input-level perturbations (Gaussian noise, blur, occlusion) during training; this would show whether the reported resilience comes from the proposed loss or from multi-objective regularization more generally.
  • The homotopy scheduling idea is not face-specific and could transfer to other structured prediction tasks where subpopulation fairness matters, such as medical image segmentation or autonomous driving perception.
  • Since the paper treats the GAN as a passive consumer of segmentation maps, an obvious next step is bidirectional optimization in which generator feedback refines the parser during training.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a multi-objective face parsing framework that combines accuracy (Dice loss), fairness (variance of per-group mIoU), and robustness (negative mIoU under perturbed outputs) via a homotopy-based weighting schedule. The resulting U-Net parsers are used to condition a Pix2PixHD GAN and a ControlNet diffusion pipeline, and the paper reports segmentation mIoU, fairness variance, and generative FID/LPIPS results. The central claim is that multi-objective training improves demographic consistency and robustness in parsing, which in turn improves GAN- and diffusion-based face synthesis.

Significance. If the claims were fully supported, the work would address a timely and important problem: making face parsers fair and robust and demonstrating downstream benefits for generative models. The paper also attempts to extend robustness beyond clean accuracy and to evaluate fairness in a downstream synthesis context. However, as presented, the evidence is undermined by a mismatch between the robustness training objective and the evaluation protocol, the use of non-differentiable losses without stated surrogates, and internal numerical inconsistencies. The topic is significant, but the current manuscript does not establish the claimed cause-and-effect relationship.

major comments (4)
  1. [Section III-D, Algorithm 1; Section IV-B] The robustness loss in Algorithm 1 (line 7-8) perturbs the network's output logits (outputs_noisy = outputs + random noise()), while the robustness evaluation in Section IV-B perturbs the input images (Gaussian noise, blur, occlusion, salt-and-pepper). This is a fundamental training/evaluation mismatch: the proposed loss provides no training signal for invariance to input corruption, so the reported robustness improvements (Figure 3, Figure 5, and Table II) cannot be attributed to the proposed method. Please either modify the training to include input-side perturbations (e.g., via a differentiable corruption layer) or restrict the claims to output-logit robustness and evaluate accordingly.
  2. [Section III-D, Algorithm 1, lines 8-9] Both Lrob = -mIoU(softmax(outputs_noisy), m) and Lfair = Var[mIoUg] depend on mIoU, which is a non-differentiable metric. The paper does not specify a differentiable surrogate or a gradient approximation for these terms. As written, the losses cannot be back-propagated, making the algorithm unimplementable. Please state the exact surrogate (e.g., soft IoU, per-pixel cross-entropy weighted by class, or a straight-through estimator) and justify its use.
  3. [Table III] The class-wise mIoU values in Table III are implausible: within each model, all 19 columns are reported as the same value (e.g., 73.87 for Single Objective in columns 0-5 and 7-18), which cannot be correct for genuine per-class IoU. Furthermore, Table I reports Multi-Objective (Alt. Fairness) mIoU as 73.81%, while Table III reports 74.21% for the same model. These inconsistencies call into question the fairness evaluation and the claim that the multi-objective model outperforms the single-objective model across all classes. Please verify the raw per-class results and correct the tables.
  4. [Section IV-B, Table II] The text of Section IV-B says robustness is measured by mIoU degradation under increasing severity, but Table II reports FID and LPIPS values for GAN outputs, not segmentation mIoU. The only segmentation-level robustness results are shown in Figure 5, but no quantitative table or statistical summary accompanies them. Please align the textual description with the actual metrics and provide a table of mIoU (with standard deviations) for each perturbation and severity level.
minor comments (5)
  1. [Section III-E] The homotopy schedules are described only qualitatively (linear, sigmoid, piecewise). Please provide the explicit functional form of h(t) (e.g., equations for α(t), β(t), γ(t)) and the chosen hyperparameters, as this is central to reproducing the method.
  2. [Section IV-D, Table IV] No statistical significance tests or multiple-seed averages are reported for the FID/LPIPS differences. Given that the GAN is trained once on maps from a particular parser, the observed differences (e.g., FID 99.93 vs. 98.87) may lie within training noise. Please report mean±std over at least three seeds or perform significance testing.
  3. [Figure 4] The plots for the two fairness-loss strategies are difficult to read because of overlapping labels, similar line styles, and unclear axis annotations. Please improve the figure clarity and add a legend that distinguishes the single-objective and multi-objective variants explicitly.
  4. [Algorithm 1, line 7] The perturbation is only described as "random noise()" without distribution, magnitude, or schedule. Please specify the noise type and severity, and state whether it is added during both forward and backward passes.
  5. [General notation] The paper uses mIoU and mIoUg without formally defining subscript g in the algorithm; please define all notation for per-group mIoU and the variance computation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper directly optimizes its stated objectives and evaluates them on held-out data.

full rationale

The paper proposes a multi-objective U-Net trained with a weighted sum of accuracy, fairness, and robustness losses, and then evaluates segmentation and downstream GAN/diffusion outputs on held-out test data. The fairness loss is defined as variance of group-wise mIoU, and fairness is reported using that same variance metric; this is direct optimization of a stated objective on training data plus evaluation on test data, not a prediction that reduces to an input by construction. The robustness loss in Algorithm 1 perturbs output logits, while Section IV-B evaluates input perturbations such as Gaussian noise, blur, and occlusion; this is a train/evaluation mismatch that weakens the robustness claim, but it is not circular reasoning. No load-bearing self-citations appear in the reference list. The ControlNet experiment is explicitly preliminary, limited to one epoch of fine-tuning, and the paper acknowledges limitations including dataset imbalance and treating GANs as passive consumers of segmentation maps. Overall, the derivation chain is empirical rather than definitional, so no circular step is present.

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

The central claim rests on hand-chosen schedule parameters and on assumptions that the fairness and robustness proxies are valid training signals. The most important issue is the robustness loss, which is defined on output noise rather than input perturbations, making the training signal mismatched with the evaluation.

free parameters (3)
  • Homotopy schedule parameters = Not specified
    The linear, sigmoid, and piecewise weight schedules for alpha, beta, gamma are chosen by hand without ablation or sensitivity analysis. Exact equations and parameter values are not given.
  • Perturbation severity levels = 0.1, 0.3, 0.5
    Severity levels for Gaussian, blur, occlusion, and salt-and-pepper are selected without justification and affect the robustness numbers reported.
  • Training hyperparameters = lr=1e-4, Adam, 30 epochs (from Figure 2)
    Standard choices but no batch size or schedule details are provided; these affect all reported metrics.
assumptions (5)
  • domain assumption CelebAMask-HQ ground truth masks and demographic attribute labels are accurate enough to train and evaluate fairness.
    The paper relies on these annotations to compute mIoU and per-group variance.
  • domain assumption mIoU variance across demographic groups is a valid and sufficient measure of fairness.
    Fairness loss and evaluation are both based on Var(mIoU_g); no justification links this metric to meaningful fairness.
  • ad hoc to paper The negative mIoU of noisy logits is a valid differentiable training signal for robustness.
    mIoU is non-differentiable; the paper does not specify a soft approximation, and the noise is applied to outputs, not inputs.
  • domain assumption ImageNet pretraining on ResNet-34 improves face parsing and transfers to fairness and robustness objectives.
    Standard practice, but no analysis of its effect on the multi-objective results is provided.
  • ad hoc to paper A one-epoch ControlNet fine-tune is sufficient to compare segmentation map quality.
    The paper admits this is preliminary; the tiny FID and LPIPS differences are not meaningful evidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Fair and Robust Face Parsing for Generative AI: A Multi-Objective Approach." pith.science (2026). https://pith.science/paper/LONPMS3C

@misc{pith2026250204391,
  author       = {Pith},
  title        = {Pith review of: Towards Fair and Robust Face Parsing for Generative AI: A Multi-Objective Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LONPMS3C}},
  note         = {Machine review of arXiv:2502.04391}
}
read the original abstract

Face parsing is a fundamental task in computer vision, enabling applications such as identity verification, facial editing, and controllable image synthesis. However, existing face parsing models often lack fairness and robustness, leading to biased segmentation across demographic groups and errors under occlusions, noise, and domain shifts. These limitations affect downstream face synthesis, where segmentation biases can degrade generative model outputs. We propose a multi-objective learning framework that optimizes accuracy, fairness, and robustness in face parsing. Our approach introduces a homotopy-based loss function that dynamically adjusts the importance of these objectives during training. To evaluate its impact, we compare multi-objective and single-objective U-Net models in a GAN-based face synthesis pipeline (Pix2PixHD). Our results show that fairness-aware and robust segmentation improves photorealism and consistency in face generation. Additionally, we conduct preliminary experiments using ControlNet, a structured conditioning model for diffusion-based synthesis, to explore how segmentation quality influences guided image generation. Our findings demonstrate that multi-objective face parsing improves demographic consistency and robustness, leading to higher-quality GAN-based synthesis.

Figures

Figures reproduced from arXiv: 2502.04391 by the authors.

Figure 1
Figure 1. Overview of Our Multi-Objective Face Parsing and Synthesis Framework. Our proposed homotopy-based multi-objective learning framework optimizes accuracy (Lacc), robustness (Lrob), and fairness (Lfair). This framework produces fairness-aware and robust segmentation maps, which are used to train two generative pipelines: (1) a GAN-based synthesis model (Pix2PixHD), where improved segmentation enhances photorealism and … view at source ↗
Figure 2
Figure 2. Comparison of α, β, and γ schedules across three homotopy methods (Linear, Sigmoid, and Piecewise) over 30 epochs. Each subplot illustrates the evolution of a parameter (α, β, or γ) as it adapts during training, highlighting the differences in transition dynamics across homotopy strategies. The legend below the figure identifies the homotopy method for each curve. • Piecewise: Abrupt changes in weight distribution a… view at source ↗
Figure 3
Figure 3. Qualitative comparison of Single-Objective and Multi-Objective models under perturbations. Blur (severity = 0.3), Gaussian Noise (severity = 0.1), and Occlusion (severity = 0.5) are applied to input images (first column). The Single-Objective model produces fragmented and inaccurate segmentations, especially in occluded and blurred regions. In contrast, the Multi-Objective model exhibits greater robustness, preservi… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Comparison of Fairness Loss Strategies on High-Disparity Demographics. The left plot represents the fairness variance-based approach, which minimizes the variance of per-group mIoU scores, indirectly reducing fairness gaps across demographic attributes. The right plot …
Figure 5
Figure 5. Figure 5: Performance comparison of mIoU across methods and pertur￾bation types under varying severities. The plot illustrates the sensitivity of Single Objective and Multi-Objective methods (Linear, Sigmoid, and Piecewise) to perturbations, categorized by Gaussian noise, blur, …
Figure 6
Figure 6. Figure 6: Impact of Segmentation Maps on GAN-Based Face Synthesis. Segmentation maps from a Single-Objective U-Net and Multi-Objective U-Nets (Linear, Sigmoid, Piecewise) serve as inputs to a Pix2Pix GAN. Single-objective segmentation introduces inconsistencies, distorting facia…
Figure 7
Figure 7. Figure 7: Effect of Segmentation Quality on Diffusion-Based Synthesis After One Epoch of Fine-Tuning. The top row shows images generated using segmentation maps from the Single-Objective U-Net, while the bottom row corresponds to the Multi-Objective (Linear) U-Net. Despite being…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 30 canonical work pages

  1. [1]

    Albuquerque, J

    I. Albuquerque, J. Monteiro, T. Doan, B. Considine, T. Falk, and I. Mitliagkas. Multi-objective training of generative adversarial net- works with multiple discriminators. In International Conference on Machine Learning, pages 202–211. PMLR, 2019

  2. [2]

    Buolamwini and T

    J. Buolamwini and T. Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Conference on fairness, accountability and transparency, pages 77–91. PMLR, 2018

  3. [3]

    L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence , 40(4):834–848, 2017

  4. [4]

    Chien, H

    C.-H. Chien, H. Fan, A. Abdelfattah, E. Tsigaridas, S. Tomov, and B. Kimia. Gpu-based homotopy continuation for minimal problems in computer vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15765–15776, 2022

  5. [5]

    C. A. Corneanu, M. O. Sim ´on, J. F. Cohn, and S. E. Guerrero. Survey on rgb, 3d, thermal, and multimodal approaches for facial expression recognition: History, trends, and affect-related applications. IEEE transactions on pattern analysis and machine intelligence , 38(8):1548–1568, 2016

  6. [6]

    P. Dhar, J. Gleason, A. Roy, C. D. Castillo, and R. Chellappa. Pass: protected attribute suppression system for mitigating bias in face recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 15087–15096, 2021

  7. [7]

    Friedrich, M

    F. Friedrich, M. Brack, L. Struppek, D. Hintersdorf, P. Schramowski, S. Luccioni, and K. Kersting. Fair diffusion: Instructing text-to-image generation models on fairness. arXiv preprint arXiv:2302.10893, 2023

  8. [8]

    Geirhos, P

    R. Geirhos, P. Rubisch, C. Michaelis, M. Bethge, F. A. Wichmann, and W. Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018

Show all 41 references
  1. [9]

    Ghiasi, Y

    G. Ghiasi, Y . Yang, D. Ramanan, and C. C. Fowlkes. Parsing occluded people. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2401–2408, 2014

  2. [10]

    Goodfellow, J

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio. Generative adversarial nets. Advances in neural information processing systems , 27, 2014

  3. [11]

    Grother, M

    P. Grother, M. Ngan, and K. Hanaoka. Face recognition vendor test (fvrt): Part 3, demographic effects. National Institute of Standards and Technology Gaithersburg, MD, 2019

  4. [12]

    Hendrycks and T

    D. Hendrycks and T. Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019

  5. [13]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems , 33:6840–6851, 2020

  6. [14]

    Karras, S

    T. Karras, S. Laine, M. Aittala, J. Hellsten, J. Lehtinen, and T. Aila. Analyzing and improving the image quality of stylegan. In Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110–8119, 2020

  7. [15]

    Kendall, Y

    A. Kendall, Y . Gal, and R. Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7482–7491, 2018

  8. [16]

    C.-H. Lee, Z. Liu, L. Wu, and P. Luo. Maskgan: Towards diverse and interactive facial image manipulation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2020

  9. [17]

    H.-M. Lee. Implementing augmented reality by using face detection, recognition and motion tracking. Journal of the Korea society of computer and information , 17(1):97–104, 2012

  10. [18]

    J. Lin, Y . Yuan, T. Shao, and K. Zhou. Towards high-fidelity 3d face reconstruction from in-the-wild images using graph convolutional networks. In Proceedings of the ieee/cvf conference on computer vision and pattern recognition , pages 5891–5900, 2020

  11. [19]

    P. Luo, X. Wang, and X. Tang. Hierarchical face parsing via deep learning. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2480–2487. IEEE, 2012

  12. [20]

    Menon, A

    S. Menon, A. Damian, S. Hu, N. Ravi, and C. Rudin. Pulse: Self- supervised photo upsampling via latent space exploration of generative models. In Proceedings of the ieee/cvf conference on computer vision and pattern recognition , pages 2437–2445, 2020

  13. [21]

    Merler, N

    M. Merler, N. Ratha, R. S. Feris, and J. R. Smith. Diversity in faces. arXiv preprint arXiv:1901.10436 , 2019

  14. [22]

    Minaee, Y

    S. Minaee, Y . Boykov, F. Porikli, A. Plaza, N. Kehtarnavaz, and D. Terzopoulos. Image segmentation using deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence , 44(7):3523–3542, 2021

  15. [23]

    Morgan and A

    A. Morgan and A. Sommese. Computing all solutions to polynomial systems using homotopy continuation. Applied Mathematics and Computation, 24(2):115–138, 1987

  16. [24]

    Navon, I

    A. Navon, I. Achituve, H. Maron, G. Chechik, and E. Fetaya. Auxiliary learning by implicit differentiation. arXiv preprint arXiv:2007.02693, 2020

  17. [25]

    S. Park, J. Lee, P. Lee, S. Hwang, D. Kim, and H. Byun. Fair con- trastive learning for facial attribute classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 10389–10398, 2022

  18. [26]

    Park, M.-Y

    T. Park, M.-Y . Liu, T.-C. Wang, and J.-Y . Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 2337–2346, 2019

  19. [27]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

  20. [28]

    Sener and V

    O. Sener and V . Koltun. Multi-task learning as multi-objective optimization. Advances in neural information processing systems , 31, 2018

  21. [29]

    Sharma and V

    S. Sharma and V . Kumar. A comprehensive review on multi- objective optimization techniques: Past, present and future. Archives of Computational Methods in Engineering , 29(7):5605–5633, 2022

  22. [30]

    Standley, A

    T. Standley, A. Zamir, D. Chen, L. Guibas, J. Malik, and S. Savarese. Which tasks should be learned together in multi-task learning? In International conference on machine learning , pages 9120–9132. PMLR, 2020

  23. [31]

    C. H. Sudre, W. Li, T. Vercauteren, S. Ourselin, and M. Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third Interna...

  24. [32]

    S. Tan, Y . Shen, and B. Zhou. Improving the fairness of deep gen- erative models without retraining. arXiv preprint arXiv:2012.04842 , 2020

  25. [33]

    C. Wang, C. Xu, X. Yao, and D. Tao. Evolutionary generative ad- versarial networks. IEEE Transactions on Evolutionary Computation , 23(6):921–934, 2019

  26. [34]

    Wang and W

    M. Wang and W. Deng. Deep face recognition: A survey. Neurocom- puting, 429:215–244, 2021

  27. [35]

    Wang, M.-Y

    T.-C. Wang, M.-Y . Liu, J.-Y . Zhu, A. Tao, J. Kautz, and B. Catanzaro. High-resolution image synthesis and semantic manipulation with con- ditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 8798–8807, 2018

  28. [36]

    Z. Wu, D. Lischinski, and E. Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12863–12872, 2021

  29. [37]

    Z. Xu, X. Yu, Z. Hong, Z. Zhu, J. Han, J. Liu, E. Ding, and X. Bai. Facecontroller: Controllable attribute editing for face in the wild. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 35, pages 3083–3091, 2021

  30. [38]

    Y . Yu, W. Zhang, and Y . Deng. Frechet inception distance (fid) for evaluating gans. China University of Mining Technology Beijing Graduate School, 3, 2021

  31. [39]

    Zhang, A

    L. Zhang, A. Rao, and M. 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

  32. [40]

    Zhang, P

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018

  33. [41]

    J.-Y . Zhu, T. Park, P. Isola, and A. A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision , pages 2223– 2232, 2017

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.