Pith. sign in

REVIEW 3 major objections 5 minor 27 references

TSCL:Multi-party loss Balancing scheme for deep learning Image steganography based on Curriculum learning

T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read The paper argues that its two-stage curriculum-learning loss scheduler TSCL improves steganography quality, decoding accuracy, and steganalysis resistance on ALASKA2, VOC2012, and ImageNet.

desk verdict The method is a sensible, clearly described mash-up of known loss-balancing tricks, but the paper's own tables refute its headline claim that security improves, so it needs major revision before it can be taken seriously. read the letter →

arxiv 2504.18348 v1 pith:A6GNWMHY submitted 2025-04-25 cs.CV cs.AIcs.CR

Fengchun Liu. Tong Zhang , Chunying Zhang This is my paper · ORCID
classification cs.CVcs.AIcs.CR
keywords imagesteganographycurriculumlearninglossbalancingmulti-taskadversarialtrainingsteganalysisdeepscheduling
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

Deep-learning image steganography trains an encoder, a decoder, and a steganalysis detector together, and previous work usually sums the embedding, recovery, and steganalysis losses with fixed weights. This paper argues that fixed weights ignore both the model's changing learning state and the natural priority of the three tasks: the hidden image must first be invisible, second be recoverable, and third resist detection. It proposes TSCL, a two-stage loss scheduler: first, a curriculum raises the weight of one loss at a time in the order embedding, decoding, steganalysis; second, each weight is adjusted by the ratio of that loss between consecutive iterations, with prior coefficients keeping embedding dominant. The paper reports that on ALASKA2, VOC2012, and ImageNet, TSCL improves steganographic image quality and decoding accuracy, and on several settings also improves resistance to steganalysis, relative to fixed-weight training.

What carries the argument

The load-bearing object is the TSCL scheduler, which wraps the existing total loss $L_{\text{total}}=L_{\text{Encode}}+L_{\text{Decode}}+L_{\text{steganalysis}}$ with a two-phase weighting scheme. Phase one, the prior curriculum, uses a scheduling function or fixed step rule to raise the weight of the embedding loss first, then the decoding loss, then the steganalysis loss, so the model's focus moves through the three tasks in that order. Phase two, loss dynamics control, measures each task's learning speed at iteration $t$ by the loss-decline ratio $W_k(t)=L_k(t-1)/L_k(t-2)$ and sets the final weight to $\lambda_k(t)=D_k W_k(t)$, with prior coefficients $D_{\text{Encode}}=1$, $D_{\text{Decode}}=0.8$, and $D_{\text{steganalysis}}=0.4$. This two-stage mechanism converts a fixed-weight steganography trainer into one whose loss emphasis follows both a designed curriculum and the measured difficulty of each task.

What would settle it

Train the same model on ALASKA2 at 3 bpp with the curriculum order reversed—steganalysis first, then decoding, then embedding—keeping every other TSCL setting identical; if the reversed order matches or beats the proposed one on PSNR, decoding accuracy, and steganalysis score, the claimed ordering is not the source of the reported gains.

Watch

Extended reading notes

Core claim

The paper's central claim is that in an encoder-decoder-steganalysis steganography framework, the schedule of loss weights is itself a control variable, and a curriculum order that matches task priority—embedding first, recovery second, resistance to detection last—produces better final models than static weighting. TSCL implements this in two stages. In the curriculum-control stage, weights are raised one task at a time either by a continuous function (sine, linear, or exponential) or by discrete step changes, so the model masters invisibility before moving to decoding accuracy and then to steganalysis resistance. In the loss-dynamics stage, the scheduler computes the per-task loss-decline ratio $W_k(t)=L_k(t-1)/L_k(t-2)$ and sets the effective weight to $\lambda_k(t)=D_k W_k(t)$, where the prior coefficients satisfy $D_{\text{Encode}} > D_{\text{Decode}} > D_{\text{steganalysis}}$. The paper reports gains in PSNR, SSIM, and decoding accuracy on ALASKA2, VOC2012, and ImageNet, and states the overall result as simultaneous improvement in steganography quality, decoding accuracy, and security.

Load-bearing premise

The load-bearing premise is that the fixed task ordering in TSCL is right—embedding loss should stay dominant, then decoding, then steganalysis—and that this priority holds across datasets, capacities, and architectures; if the ideal ordering differs, the scheduler's benefits could vanish or reverse.

Editorial extensions

If this is right

  • If the paper's claim is right, any encoder-decoder-steganalysis steganography trainer can adopt TSCL by replacing fixed loss weights with the two-stage scheduler, leaving the network unchanged.
  • The embedding-first curriculum means early training effort on hiding quality is preserved when later phases shift focus to decoding and security, because previous loss weights are not lowered.
  • Since the second stage uses only scalar loss values and no gradients, TSCL adds negligible memory and compute overhead, making it a practical drop-in for large-scale training.
  • The largest decoding-accuracy gains in the reported tables appear at 3 bpp capacity, suggesting TSCL helps most when the decoding task is hardest.

Reading between the lines

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

  • Because TSCL only reweights losses that already exist in an encoder-decoder-steganalysis pipeline, it should transfer to invertible-network steganography models without changing the network, a direction the paper does not test.
  • The ImageNet security decrease shown in the tables suggests the fixed order $D_{\text{Encode}} > D_{\text{Decode}} > D_{\text{steganalysis}}$ may encode dataset-specific task difficulty; an adaptive variant that learns the order from loss-decline rates is a plausible next step.
  • The loss-decline ratio $W_k(t)$ is noisy from iteration to iteration, especially late in training; smoothing it over a short window could make the second stage more stable while preserving the mechanism.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 TSCL, a two-phase loss-weight scheduler for deep-learning image steganography models that combine an embedding network, a decoding network, and a steganalysis network. The first phase, "curriculum control," sequentially up-weights the embedding loss, then the decoding loss, then the steganalysis loss; the second phase, "loss dynamics control," adjusts weights according to the ratio of each loss between successive iterations, multiplied by a priori coefficients that keep the embedding loss dominant. The authors report experiments on ALASKA2, VOC2012, and ImageNet at capacities 1-3 bpp and claim simultaneous improvements in steganographic image quality, decoding accuracy, and security.

Significance. If the central claim were supported, TSCL would be a simple, drop-in loss-balancing scheme for a widely used class of steganography models, and the paper's use of three datasets and several curriculum-schedule variants would provide a reasonable experimental skeleton. However, the paper's own tables contradict the headline claim: security (measured by the steganalysis score, lower is better) worsens in 6 of 9 configurations in Table 1, including all ImageNet rows, and the text itself concedes that security decreased on ImageNet. In addition, key schedule parameters are unspecified and the a priori coefficients in Eq. (12) are hand-set to keep embedding loss dominant, which pre-configures one of the claimed benefits. The contribution is therefore not established as stated.

major comments (3)
  1. [§4.3.1, Table 1; Abstract; §5] The abstract and conclusion claim that TSCL "improves the quality of steganography, decoding accuracy and security," but Table 1 shows that the steganalysis score (lower is better) worsens under TSCL in 6 of 9 configurations: ALASKA2 2 bpp (0.402 to 0.442), VOC2012 1 bpp (0.434 to 0.457), VOC2012 3 bpp (0.357 to 0.449), and all three ImageNet rows (0.362 to 0.414, 0.325 to 0.391, and 0.413 to 0.423). Table 4 likewise shows VOC2012 security worsening from 0.357 to 0.449 under the full TSCL scheme. Since one of the three headline metrics is not improved by the method in most reported configurations, the central claim as stated is internally inconsistent and unsupported by the paper's own evidence.
  2. [§3.3, Eq. (12); §4.2] The a priori coefficients D_Encode=1, D_Decode=0.8, D_steganalysis=0.4 are hand-set to keep the embedding loss dominant, and the second-stage weight update in Eq. (13) multiplies each loss-decline ratio by these coefficients. Consequently, the observation that imperceptibility does not degrade is partly guaranteed by construction, while the steganalysis loss is always the least weighted, which makes a security improvement structurally less likely. The paper provides no sensitivity analysis with respect to this ordering, and the ImageNet results suggest the ordering may be dataset-dependent. A concrete ablation varying the relative values and order of D_Encode, D_Decode, and D_steganalysis is needed before the claimed benefits can be attributed to the proposed method.
  3. [§3.2, §4.2] The curriculum scheduler is not reproducible from the manuscript because the schedule hyperparameters a0, a1, a2, C1, C2, C3, the iteration-step sizes, and the exact α values used in the experiments are never reported. Equations (8)-(10) define families of schedule functions, but no values are given for any dataset or capacity, and the text states only that the maximum number of iterations is 120 without connecting this to the epoch-based curriculum boundaries. Since the curriculum timing is the core mechanism of TSCL, the reported comparisons in Tables 2-4 cannot be reproduced or independently evaluated without these parameters.
minor comments (5)
  1. [§4.1] The dataset description contains a typo: "the datasets used are three large public datasets, ImageNet, ALASKA2 and ImageNet," where VOC2012 is presumably meant in one of the two mentions.
  2. [§2.1, References] The text cites "Li et al [17]" for a new low-frequency wavelet loss and FSIM loss, but reference [17] is Chekatamala et al.; the citation-to-reference mapping is inconsistent.
  3. [Figures 2-4] Figure captions and labels alternate between "TSLS" and "TSCL" (e.g., "Cover TSLS Baseline" and "TSLSOnly_loss"), which is confusing and should be corrected.
  4. [§4.3, Tables 1-4] All experimental comparisons report a single run with no error bars, confidence intervals, or significance tests; given the small differences in some metrics (e.g., accuracy 0.99 vs. 1.00), the robustness of the reported improvements is unclear.
  5. [§3.2, Eq. (10)] The exponential schedule function is not properly normalized: for epoch = C2 the argument (e^{(epoch-C1)/(C2-C1)} - 1) evaluates to e-1 rather than 1, so the claimed transition to a2 at epoch = C2 is not smooth; please clarify the intended formula.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TSCL's a priori weights are explicit method inputs and the reported improvements are empirical; the paper's security claim is internally contradicted by its own Table 1, but that is a correctness issue, not circularity.

full rationale

TSCL's two stages are explicit training heuristics, not quantities fitted to the evaluation metrics: Eq. (8)-(10) define the curriculum schedules, Eq. (12) fixes the a priori coefficients (D_Encode=1, D_Decode=0.8, D_steganalysis=0.4) before training, and Eq. (11) is a deterministic ratio of successive loss values. The paper then measures the resulting model on held-out test sets against a fixed-weight baseline, so no reported success metric is used to fit the method's parameters; the derivation chain is self-contained and the claimed improvements are empirical, not tautological. Although the embedding loss is given the largest weight by construction, Table 1 shows the outcomes are not forced: several metrics worsen under TSCL (e.g., ALASKA2 2bpp SSIM 0.99547 to 0.99434 and steganalysis 0.402 to 0.442; VOC2012 3bpp steganalysis 0.357 to 0.449; ImageNet steganalysis worsens in all three rows). The paper's abstract and conclusion claim security improvement, yet its own Table 1 shows steganalysis scores worsen in 5 of 9 configurations, and Section 4.3.1 concedes 'security ... slightly decreased on the ImageNet dataset'; this is an internal consistency/correctness problem, not a circularity. There are no load-bearing self-citations: the author-name coincidences in the reference list refer to different researchers, and no uniqueness theorem or prior result by the present authors is invoked to force the chosen ordering. Therefore the circularity score is 0.

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

The central claim depends on several hand-set scheduler parameters and on the assumed task ordering. The most consequential free parameters are the unstated curriculum epochs C1, C2, C3 and the schedule amplitudes a0, a1, a2, which directly control the loss weights. The prior coefficients D_k and the encoding loss composition weights are also chosen by hand, and the paper does not test sensitivity to these choices. No new physical or conceptual entities are introduced.

free parameters (4)
  • a0, a1, a2 (schedule amplitudes for each loss) = not reported
    Used in Eqs. 8-10 to control loss weight growth, but concrete values for the experiments are never stated.
  • C1, C2, C3 (curriculum epoch boundaries) = not reported
    Define when each loss weight changes in the curriculum stage; no values are given in Section 4.2 or elsewhere.
  • D_Encode, D_Decode, D_steganalysis = 1, 0.8, 0.4
    Hand-set prior coefficients in Eq. 12 that keep encoding loss dominant; their values are chosen by the authors, not derived or tuned through ablation.
  • Encoding loss composition weights = SSIM 0.5, MSSSIM 0.5, RMSE 0.3
    Reported in Section 4.2 to combine three similarity metrics into the encoding loss; no sensitivity analysis is shown.
assumptions (4)
  • domain assumption The three steganography tasks (embedding, decoding, steganalysis) can be optimized by linearly combining their losses with time-varying weights.
    Invoked throughout Section 3; no convergence or optimality guarantee is given for a weighted sum of these adversarial losses.
  • ad hoc to paper Imperceptibility is the most important task, followed by decoding accuracy, then security, as encoded in Eq. 12.
    This ordering is asserted in Section 3.2 and baked into the prior coefficients, but no dataset-specific evidence supports it.
  • domain assumption The loss decline ratio Loss(t-1)/Loss(t-2) measures task learning difficulty and should be used to scale loss weights.
    Introduced in Eq. 11 without derivation; the paper does not justify why a larger ratio (slower decline) maps to a proportionally larger weight.
  • domain assumption The steganalysis network from Zhang et al. [16] is a valid adversarial evaluator.
    Used as the security loss component in all experiments without independent verification of its strength or calibration.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TSCL:Multi-party loss Balancing scheme for deep learning Image steganography based on Curriculum learning." pith.science (2026). https://pith.science/paper/A6GNWMHY

@misc{pith2026250418348,
  author       = {Pith},
  title        = {Pith review of: TSCL:Multi-party loss Balancing scheme for deep learning Image steganography based on Curriculum learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A6GNWMHY}},
  note         = {Machine review of arXiv:2504.18348}
}
read the original abstract

For deep learning-based image steganography frameworks, in order to ensure the invisibility and recoverability of the information embedding, the loss function usually contains several losses such as embedding loss, recovery loss and steganalysis loss. In previous research works, fixed loss weights are usually chosen for training optimization, and this setting is not linked to the importance of the steganography task itself and the training process. In this paper, we propose a Two-stage Curriculum Learning loss scheduler (TSCL) for balancing multinomial losses in deep learning image steganography algorithms. TSCL consists of two phases: a priori curriculum control and loss dynamics control. The first phase firstly focuses the model on learning the information embedding of the original image by controlling the loss weights in the multi-party adversarial training; secondly, it makes the model shift its learning focus to improving the decoding accuracy; and finally, it makes the model learn to generate a steganographic image that is resistant to steganalysis. In the second stage, the learning speed of each training task is evaluated by calculating the loss drop of the before and after iteration rounds to balance the learning of each task. Experimental results on three large public datasets, ALASKA2, VOC2012 and ImageNet, show that the proposed TSCL strategy improves the quality of steganography, decoding accuracy and security.

Figures

Figures reproduced from arXiv: 2504.18348 by the authors.

Figure 1
Figure 1. Discrete loss scheduling and continuous loss scheduling. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Comparison between TSCL scheme and steganography images using only curriculum control and only loss [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Comparison of cover image and stego image of TSCL scheme under 1-3 bpp capacity steganogayphy. [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Comparison of image steganography quality between TSCL scheme and baseline scheme model. [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 23 canonical work pages

  1. [1]

    Automatic steganographic distortion learning using a generative adversarial network

    Weixuan Tang, Shunquan Tan, Bin Li, and Jiwu Huang. Automatic steganographic distortion learning using a generative adversarial network. IEEE Signal Processing Letters, 24(10):1547–1551, 2017

    work page 2017

  2. [2]

    Acgis: Adversarial cover generator for image steganography with noise residuals features-preserving

    Junxue Yang and Xin Liao. Acgis: Adversarial cover generator for image steganography with noise residuals features-preserving. Signal Processing: Image Communication, 113:116927, 2023

    work page 2023

  3. [3]

    Hinet: Deep image hiding by invertible network

    Junpeng Jing, Xin Deng, Mai Xu, Jianyi Wang, and Zhenyu Guan. Hinet: Deep image hiding by invertible network. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4733–4742, 2021

    work page 2021

  4. [4]

    Robust invertible image steganography

    Youmin Xu, Chong Mou, Yujie Hu, Jingfen Xie, and Jian Zhang. Robust invertible image steganography. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7875–7884, 2022

    work page 2022

  5. [5]

    Generative adversarial networks for image steganogra- phy

    Denis V olkhonskiy, Boris Borisenko, and Evgeny Burnaev. Generative adversarial networks for image steganogra- phy. 2016

    work page 2016

  6. [6]

    An improved steganography without embedding based on attention gan

    Cong Yu, Donghui Hu, Shuli Zheng, Wenjie Jiang, Meng Li, and Zhong-qiu Zhao. An improved steganography without embedding based on attention gan. Peer-to-Peer Networking and Applications, 14(3):1446–1457, 2021

    work page 2021

  7. [7]

    Gan-based spatial image steganography with cross feedback mechanism

    Fengyong Li, Zongliang Yu, and Chuan Qin. Gan-based spatial image steganography with cross feedback mechanism. Signal Processing, 190:108341, 2022

    work page 2022

  8. [8]

    Gan-based image steganography for enhancing security via adversarial attack and pixel-wise deep fusion

    Chao Yuan, Hongxia Wang, Peisong He, Jie Luo, and Bin Li. Gan-based image steganography for enhancing security via adversarial attack and pixel-wise deep fusion. Multimedia Tools and Applications, 81(5):6681–6701, 2022

    work page 2022

Show all 27 references
  1. [9]

    Stegastylegan: towards generic and practical generative image steganography

    Wenkang Su, Jiangqun Ni, and Yiyan Sun. Stegastylegan: towards generic and practical generative image steganography. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 240–248, 2024

    2024

  2. [10]

    Generating steganographic images via adversarial training

    Jamie Hayes and George Danezis. Generating steganographic images via adversarial training. Advances in neural information processing systems, 30, 2017

    2017

  3. [11]

    Steganogan: High capacity image steganography with gans

    Kevin Alex Zhang, Alfredo Cuesta-Infante, Lei Xu, and Kalyan Veeramachaneni. Steganogan: High capacity image steganography with gans. arXiv preprint arXiv:1901.03892, 2019

    1901 arXiv

  4. [12]

    Image steganography with deep orthogonal fusion of multi-scale channel attention

    Yinyin Peng, Donghui Hu, Gang Pei, and Yaofei Wang. Image steganography with deep orthogonal fusion of multi-scale channel attention. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4850–4854. IEEE, 2024

    2024

  5. [13]

    High invisibility image steganography with wavelet transform and generative adversarial network

    Ye Yao, Junyu Wang, Qi Chang, Yizhi Ren, and Weizhi Meng. High invisibility image steganography with wavelet transform and generative adversarial network. Expert Systems with Applications, 249:123540, 2024

    2024

  6. [14]

    Image hide with invertible network and swin transformer

    Yuhuan Feng, Yunjie Liu, Hongjuan Wang, Jin Dong, Rujia Wang, and Chunpeng Tian. Image hide with invertible network and swin transformer. In International Conference on Data Mining and Big Data , pages 385–394. Springer, 2022

    2022

  7. [15]

    iscmis: Spatial-channel attention based deep invertible network for multi-image steganography

    Fengyong Li, Yang Sheng, Xinpeng Zhang, and Chuan Qin. iscmis: Spatial-channel attention based deep invertible network for multi-image steganography. IEEE Transactions on Multimedia, 26:3137–3152, 2023

    2023

  8. [16]

    Invisible steganography via generative adversarial networks

    Ru Zhang, Shiqi Dong, and Jianyi Liu. Invisible steganography via generative adversarial networks. Multimedia tools and applications, 78(7):8559–8575, 2019

    2019

  9. [17]

    Analysis of deep steganography robustness using various loss functions

    Vamshi Chekatamala, P Malathi, and Gireesh Kumar. Analysis of deep steganography robustness using various loss functions. In 2022 6th International Conference on Intelligent Computing and Control Systems (ICICCS), pages 148–153. IEEE, 2022. 14

    2022

  10. [18]

    Multi-task learning using uncertainty to weigh losses for scene geometry and semantics

    Alex Kendall, Yarin Gal, and Roberto 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

    2018

  11. [19]

    Dynamic task prioritization for multitask learning

    Michelle Guo, Albert Haque, De-An Huang, Serena Yeung, and Li Fei-Fei. Dynamic task prioritization for multitask learning. In Proceedings of the European conference on computer vision (ECCV), pages 270–287, 2018

    2018

  12. [20]

    Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks

    Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794–803. PMLR, 2018

    2018

  13. [21]

    End-to-end multi-task learning with attention

    Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1871–1880, 2019

    2019

  14. [22]

    Debiased contrastive curriculum learning for progressive generalizable person re-identification

    Tiantian Gong, Kaixiang Chen, Liyan Zhang, and Junsheng Wang. Debiased contrastive curriculum learning for progressive generalizable person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 33(10):5947–5958, 2023

    2023

  15. [23]

    Curriculum learning for goal-oriented semantic communications with a common language

    Mohammad Karimzadeh Farshbafan, Walid Saad, and Merouane Debbah. Curriculum learning for goal-oriented semantic communications with a common language. IEEE Transactions on Communications, 71(3):1430–1446, 2023

    2023

  16. [24]

    Recrecnet: Rectangling rectified wide- angle images by thin-plate spline model and dof-based curriculum learning

    Kang Liao, Lang Nie, Chunyu Lin, Zishuo Zheng, and Yao Zhao. Recrecnet: Rectangling rectified wide- angle images by thin-plate spline model and dof-based curriculum learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10800–10809, 2023

    2023

  17. [25]

    Curriculum learning of multiple tasks

    Anastasia Pentina, Viktoriia Sharmanska, and Christoph H Lampert. Curriculum learning of multiple tasks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5492–5500, 2015

    2015

  18. [26]

    Curriculum pre-training for end-to-end speech translation

    Chengyi Wang, Yu Wu, Shujie Liu, Ming Zhou, and Zhenglu Yang. Curriculum pre-training for end-to-end speech translation. arXiv preprint arXiv:2004.10093, 2020

    2004 arXiv

  19. [27]

    Progressive growing of gans for improved quality, stability, and variation

    Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 15

    2017 arXiv

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.