Pith. sign in

REVIEW 5 major objections 4 minor 43 references

Imbalance-Robust and Sampling-Efficient Continuous Conditional GANs via Adaptive Vicinal Learning and Auxiliary Regularization

T0 review · 5 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read CcGAN-AVAR, an adaptive-vicinity, multi-task-discriminator extension of the continuous conditional GAN, matches or surpasses the generation quality of the diffusion-based CCDM on eleven settings while sampling 300x to 2000x faster.

desk verdict Adaptive vicinity for CcGAN is a sensible, well-ablated contribution, but the paper overstates its quality claim given single-run SFID comparisons to external baselines. read the letter →

arxiv 2508.01725 v5 pith:JN3LUOMU submitted 2025-08-03 cs.LG cs.CV

Xin Ding , Yun Chen , Yongwei Wang , Kao Zhang , Sen Zhang , Peibei Cao , Xiangxue Wang This is my paper · ORCID
classification cs.LGcs.CV
keywords continuousconditionalgenerativemodelingGANadaptivevicinitydensityratioestimationf-divergencepenaltysamplingefficiencydataimbalancelabelconsistency
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

Conditional image generation in which the condition is a continuous value — an age, a rotation angle, a steering angle — has been caught between two tools: the GAN-based CcGAN, which is fast but degrades when training labels are unevenly distributed, and the diffusion-based CCDM, which is robust but needs hundreds of denoising steps per image. The paper's goal is a framework, CcGAN-AVAR, that keeps the GAN's one-step sampling while removing its imbalance weakness. It does this in two moves: an adaptive vicinity that grows the label neighborhood around each condition until a target number of samples is captured, so dense label regions receive tight, precise neighborhoods and sparse regions receive wide ones; and a multi-task discriminator that adds a label-regression branch and a density-ratio-estimation branch, whose outputs become a label-consistency penalty and an f-divergence penalty on the generator. Across four datasets and eleven settings at resolutions from 64x64 to 256x256, the paper reports generation quality comparable to or better than CCDM — including on a new imbalanced variant of RC-49 — at 300x to 2000x the inference speed. If those results hold, continuous conditional image generation at high resolution no longer requires sacrificing either distribution quality or sampling speed.

What carries the argument

The load-bearing mechanisms are the adaptive vicinity and the multi-task discriminator. An adaptive vicinity is a label-dependent weighting kernel built by growing a zero-radius interval around a randomly noised condition $y_c$ toward the nearer neighboring label until the interval contains at least $N_{AV}$ samples, giving a radius $\kappa_{y_c} = \max(\kappa_l, \kappa_r)$ and a decay rate $\nu_{y_c} = 1/\kappa_{y_c}^2$; soft adaptive weighting assigns each sample the weight $\exp(-\nu_{y_c}(y - y_c)^2)$, and hybrid adaptive weighting multiplies that by the indicator of the interval $[y_c - \kappa_{y_c}, y_c + \kappa_{y_c}]$. A single hyperparameter $N_{AV}$ thus replaces the fixed $\kappa$ and $\nu$ of the original CcGAN, and the paper gives a histogram-based heuristic for setting it. The multi-task discriminator shares one convolutional backbone across three heads: the adversarial head with label projection, a regression head predicting the condition label and trained with a $\gamma$-insensitive hinge loss, and a density-ratio head $f_{\mathrm{dre}}$ estimating $r(x|y) = p_r(x|y)/p_g(x|y)$ under a penalized softplus loss. The two auxiliary heads feed two generator regularizers — a mean-absolute-error label-consistency penalty and the f-divergence penalty $L_G^f = \mathrm{mean}((f_{\mathrm{dre}}(D_{\mathrm{conv}}(x_g, y_g)) - 1)^2)$ — which the paper identifies, via the identity $\mathrm{Div}_{\chi^2}(p_r \| p_g) = \mathbb{E}_{p_g}[(r(x|y) - 1)^2]$, with the Pearson $\chi^2$ divergence up to the estimation error of the ratio head.

What would settle it

Use an independently trained, high-capacity density-ratio estimator on fixed slices of the label space during training and compare its per-slice ratio estimates with the outputs of the paper's $f_{\mathrm{dre}}$ head; the central claim fails if the penalty $L_G^f$ decreases while the independent Pearson $\chi^2$ divergence rises, or if $f_{\mathrm{dre}}$ drifts away from the independent ratio as training proceeds. A second check targets the imbalance claim: train CcGAN-AVAR on a dataset whose label distribution has a tail so sparse that $N_{AV}$ forces the vicinity radius to span several times the local label spacing, then inspect the label score and per-condition FID in exactly that tail — the method's quality claim collapses if label consistency breaks down precisely where the vicinity is widest.

Watch

Extended reading notes

Core claim

The paper's central claim is that the fixed-size vicinity is what limits CcGANs on skewed label distributions, and that replacing it with a density-dependent vicinity plus two auxiliary discriminator signals closes the gap with diffusion models. The adaptive vicinity keeps the original vicinal discriminator loss but substitutes a radius that expands around each target condition $y_c$ until at least $N_{AV}$ training samples fall inside it; the final radius $\kappa_{y_c}$ sets a local decay rate $\nu_{y_c} = 1/\kappa_{y_c}^2$ in the soft case, and in the hybrid case all weight outside $[y_c - \kappa_{y_c}, y_c + \kappa_{y_c}]$ is truncated to zero. The authors interpret this as tuning a local bias-variance trade-off: narrow vicinities in dense regions preserve label consistency, wide vicinities in sparse regions keep training stable, and the truncation cuts objective-level mixing between distant conditions. The discriminator is restructured into three heads on one backbone — an adversarial output with label projection, a regression head trained with a $\gamma$-insensitive hinge loss, and a density-ratio head $f_{\mathrm{dre}}$ trained with a penalized softplus loss — and the generator loss gains a mean-absolute-error term on predicted labels and an f-divergence penalty, $\mathrm{mean}((f_{\mathrm{dre}} - 1)^2)$, identified with the Pearson $\chi^2$ divergence between the real and generated conditional distributions. Empirically the framework matches or surpasses CCDM's SFID and label-consistency scores while keeping sampling 300x to 2000x faster, and it is the hybrid adaptive vicinity that usually delivers the best label consistency.

Load-bearing premise

The generator's density-ratio penalty is claimed to approximate the Pearson $\chi^2$ divergence, and that approximation requires the discriminator's ratio-estimation head to stay accurate while the generator keeps changing the very distribution the head is trying to measure — a non-stationary estimation problem for which the paper supplies no error, bias, or convergence analysis.

Editorial extensions

If this is right

  • CcGAN-AVAR preserves one-step GAN sampling while matching CCDM's generation quality, so high-resolution continuous conditional generation (ages, angles, temperatures) becomes practical for interactive and real-time pipelines rather than batch offline synthesis.
  • The stability under imbalance is visible in the paper's own tables: SFID on balanced RC-49 and imbalanced RC-49-I remains nearly identical for CcGAN-AVAR, while vanilla CcGAN and Dual-NDA degrade substantially on the imbalanced variant.
  • Because $N_{AV}$ replaces the two hand-tuned vicinity parameters of CcGAN and is estimated from the label histogram of the training set, the mechanism transfers to a new dataset without per-dataset radius search.
  • The multi-task discriminator is a modular recipe: the paper demonstrates the three-headed discriminator on SNGAN, DCGAN, SAGAN, and BigGAN backbones, with the simplest backbone (SNGAN) performing best on these benchmarks.
  • The hybrid adaptive vicinity usually beats the soft variant on label consistency, indicating that truncating samples far from the condition does real work at the objective level rather than merely re-weighting them.

Reading between the lines

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

  • Testable extension: the adaptive-vicinity construction is a density-aware kernel whose logic is not confined to GANs — the same $N_{AV}$-based radius could be dropped into diffusion or flow-matching models, or into any estimator that weights training samples by condition proximity, to test whether imbalance handling transfers beyond the GAN objective.
  • Implicit implication: the density-ratio penalty inherits a non-stationary target, so one could make it stationary — for instance by feeding the generator penalty through an EMA of the ratio head — and then measure per label slice whether the penalty's gradient correlates with the true $\chi^2$ divergence; the paper asserts the approximation in Remark 3 without such a measurement.
  • Testable refinement: the bias-variance framing suggests scaling $N_{AV}$ with local sample density to hold the effective sample count per vicinity roughly constant across conditions, rather than setting it near the global average of adjacent-label counts as the paper's heuristic does.
  • Observation left implicit: at 256x256, CCDM's label score deteriorates sharply while CcGAN-AVAR's improves, suggesting diffusion's iterative refinement may be more sensitive to label-conditioning error at high resolution; a per-condition FID breakdown would show where the gain actually comes from.
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

5 major / 4 minor

Summary. The paper proposes CcGAN-AVAR, an extension of CcGAN for continuous conditional generation. It replaces fixed-size vicinities with soft/hybrid adaptive vicinities whose radius is selected from local sample density, and it augments the discriminator with regression and density-ratio-estimation branches, yielding two auxiliary generator penalties. Experiments on RC-49, a new imbalanced RC-49-I, UTKFace, and Steering Angle claim that CcGAN-AVAR matches or exceeds CCDM generation quality while being 300x--2000x faster at inference.

Significance. If the empirical claims hold, the contribution is practically valuable: it would provide a one-step continuous conditional generator with quality competitive to a diffusion baseline, along with a new imbalanced benchmark (RC-49-I) and a unified codebase. The component-wise ablations in Table III are internally consistent and show that each proposed component improves SFID or Label Score. However, the central quality comparison currently rests on single-run numbers with confounded baselines, and the promised theoretical interpretation is not present in the manuscript.

major comments (5)
  1. [§IV-A, Tables I and II] The central claim that CcGAN-AVAR 'matches or exceeds' CCDM quality is supported only by single-run SFID values with no standard deviations, no seed counts, and no statistical tests. Several decisive margins are smaller than 0.01, e.g., RC-49 (64x64) CCDM 0.049 vs CcGAN-AVAR-H 0.042, UTKFace (64x64) 0.363 vs 0.356, and RC-49-I (trimodal) 0.060 vs 0.051. SFID is an empirical estimate over finite generated samples, so sampling noise alone can plausibly exceed these differences. The authors should report mean and standard deviation over at least three seeds and, ideally, paired evaluation.
  2. [§IV-A, Table III] The comparison against CcDPM, CCDM, CcGAN, and Dual-NDA is confounded by implementation differences. The paper states that baseline numbers on RC-49, UTKFace, and Steering Angle are 'released results from [14]', produced with different code and evaluation settings, whereas CcGAN-AVAR is run in a new codebase with EMA, mixed precision, and a different backbone. Table III shows that the strong CcGAN baseline alone (without adaptive vicinity or auxiliary regularization) improves SFID on RC-49-I (bimodal) from 0.205 (vanilla CcGAN) to 0.062, a change comparable to the full method's 0.050. The reported advantage over baselines is therefore not attributable solely to the proposed components. The authors should re-run baselines in the same codebase with matched settings, or at least isolate each contribution against a matched baseline.
  3. [§III-C, Eq. (10), Remark 3] The f-divergence penalty is claimed to approximate the Pearson chi-square divergence between pr(x|y) and pg(x|y). This requires that fdre(Dconv(x,y)) estimates the true conditional density ratio r(x|y)=pr(x|y)/pg(x|y). However, Dconv and fdre are trained jointly while the generator distribution changes during training, making the density-ratio target non-stationary. No bias, variance, or convergence analysis is provided, and Remark 3 merely restates the definition of f-divergence. Either the authors should supply a formal statement with testable conditions, or they should present an empirical validation of the estimated ratio and soften the claim to a heuristic penalty.
  4. [§IV-A vs §III-B Remark 2 vs Table S.5] There is a reproducibility-critical inconsistency in the reported values of NAV. Section IV-A states NAV is set to 50 for RC-49, 30 for RC-49-I, 400 for UTKFace, and 20 for Steering Angle. Table S.5 lists NAV=50 for all RC-49-I rows. Remark 2 states that for bimodal RC-49-I, where S-bar is about 25.279, NAV was set to 30. These three statements cannot all be correct. Since NAV controls the vicinity width and the ablation in Fig. 9 shows sensitivity to NAV, the authors must reconcile the text, Remark 2, and Table S.5, and report the exact value used for each experiment.
  5. [Abstract and Introduction] The abstract promises 'a theoretical interpretation characterizing how adaptive vicinal weighting affects the local bias-variance behavior of the discriminator target, how hybrid truncation reduces objective-level cross-condition mixing, and how the density-ratio-based generator penalty approximates a Pearson Chi-square discrepancy'. No such analysis appears in the manuscript. Remarks 1--4 are algorithmic descriptions, selection heuristics, and definitional restatements. The authors should either add the promised theoretical analysis or explicitly remove this claim from the contributions.
minor comments (4)
  1. [§III-C, Remark 4] Remark 4 refers to 'a custom gamma-insensitive hinge loss (Eq. (6))', but Eq. (6) is the exponential decay weight curve, not the regression loss. The regression loss appears later without an equation number; the cross-reference should be corrected.
  2. [Table S.5] Table S.5 uses the symbol lambda_G_dre, which is not defined in Section III-C. The method defines lambda_D_dre for the discriminator's DRE loss and lambda_G_f for the generator's f-divergence penalty. The table should use the defined notation or introduce and explain lambda_G_dre.
  3. [§III-C] There is a typo in the text defining the gamma-insensitive hinge loss: 'maximum raidus' should read 'maximum radius'. Also, the sentence 'as outlined blow' in Section IV-C should read 'as outlined below'.
  4. [§IV-A] In the description of RC-49-I construction, the paper uses 'int(49 * exp(-pi * d_i))' where pi is described as a decay rate. The symbol pi could be confused with the constant 3.14159; a different symbol such as rho would improve clarity.

Circularity Check

1 steps flagged · score 3.0 of 10

One definitional step: the f-divergence penalty's 'approximation' of Pearson chi-square is true by construction; the central empirical claims are independent of this tautology.

  1. self definitional [Section III-C(2), Remark 3; Eqs. (10)-(11)]
    "Since ˆr(x|y) =fdre(Dconv(x, y))in Eq. (9) is trained during the discriminator updates to estimate the conditional density ratio r(x|y) =p r(x|y)/pg(x|y), the f-divergence penalty defined in Eq. (10) approximates Eq. (11), and minimizing LG f encourages the generator to better match the target data distribution."

    Eq. (10) defines LG_f as the sample average of (fdre(...)-1)^2, and Eq. (11) defines Pearson chi-square as E[(r-1)^2]. The claimed approximation is therefore exact by construction whenever fdre equals r; no separate estimation, bias, or convergence analysis is supplied. Remark 3 restates the definition of the f-divergence and asserts the equivalence because the network output is trained to estimate r. This is a plug-in estimator identity, not a derived result, and the paper's own abstract hedges it as holding 'up to the estimation error of the density-ratio branch.' The step is definitional rather than an independent theoretical prediction.

full rationale

The central claim of the paper is empirical: CcGAN-AVAR matches or exceeds CCDM generation quality while being 300x-2000x faster. That claim is supported by SFID, NIQE, Diversity, and Label Score measurements, with baseline numbers taken from the authors' earlier released results. Using one's own published empirical results as baselines is normal self-citation, not circularity, because those numbers are externally falsifiable measurements rather than consequences of the present paper's assumptions. The adaptive vicinity construction is a heuristic with ablation studies rather than a conclusion derived from itself. The regression penalty directly penalizes label-prediction error, which aligns with the Label Score evaluation metric, but this is standard objective-vs-metric alignment rather than a fitted parameter being renamed as a prediction. The one genuinely circular step is the f-divergence penalty's theoretical justification: Eq. (10) is defined as the plug-in Pearson chi-square estimate, so the remark that it 'approximates' Eq. (11) is true by construction if the DRE branch is accurate. This is a tautology presented as a theoretical interpretation, but it is ancillary to the main empirical comparison. Concerns about single-run SFID values without error bars and about comparing numbers produced by different codebases are evidence-quality risks, not circularity. Accordingly, the overall circularity score is low: one definitional step in a supporting theoretical remark, while the central contribution retains independent empirical content.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method depends on one dataset-tuned free parameter (NAV), a loss weight λD_dre, and several unproven assumptions about the validity of vicinal estimation and density-ratio convergence. No new physical or conceptual entities are introduced.

free parameters (2)
  • NAV (minimum effective sample count per vicinity) = RC-49: 50; RC-49-I: 30 in Section IV-A but 50 in Table S.5; UTKFace: 400; Steering Angle: 20
    NAV directly controls the adaptive vicinity radius and weight decay rate. The paper selects it per dataset using a heuristic in Remark 2, and the ablation in Fig. 9 shows performance is sensitive to it.
  • λD_dre (DRE branch loss weight) = 1.0 for RC-49 and for hybrid RC-49-I; 0.5 otherwise
    Chosen per setting without a sensitivity study; affects the balance between the density-ratio estimation and the adversarial objectives.
assumptions (4)
  • standard math Definition of f-divergence and Pearson chi-square divergence (Remark 3).
    The f-divergence penalty derivation relies on the standard identity Div_f = E_pg[f(pr/pg)] and the quadratic form f(t)=(t-1)^2.
  • domain assumption Training samples within an adaptive label vicinity provide a valid estimate of p_r(x|y_c), even when the vicinity radius is large in sparse label regions.
    Section III-B constructs the vicinity empirically from available labels; the paper does not prove that this weighted empirical estimate converges to the true conditional distribution.
  • ad hoc to paper The heuristic in Remark 2 for choosing NAV yields near-optimal performance across datasets.
    The heuristic averages adjacent-label counts and then applies manual adjustments (e.g., Steering Angle NAV=20 because many Si cluster around 2); no principled selection rule is derived.
  • ad hoc to paper The DRE branch fdre converges to the true conditional density ratio despite a non-stationary generator and joint training of the shared backbone.
    Section III-C asserts that fdre 'is trained during the discriminator updates to estimate the conditional density ratio' and uses this to justify Eq. (10), but no convergence or error bound is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Imbalance-Robust and Sampling-Efficient Continuous Conditional GANs via Adaptive Vicinal Learning and Auxiliary Regularization." pith.science (2026). https://pith.science/paper/JN3LUOMU

@misc{pith2026250801725,
  author       = {Pith},
  title        = {Pith review of: Imbalance-Robust and Sampling-Efficient Continuous Conditional GANs via Adaptive Vicinal Learning and Auxiliary Regularization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JN3LUOMU}},
  note         = {Machine review of arXiv:2508.01725}
}
read the original abstract

Recent advances in continuous conditional generative modeling, including Continuous conditional Generative Adversarial Network (CcGAN) and Continuous Conditional Diffusion Model (CCDM), estimate high-dimensional data distributions conditioned on scalar regression labels such as angles, ages, or temperatures. However, fixed-size vicinal training in CcGAN can be sensitive to non-uniform label densities, whereas CCDM relies on computationally expensive iterative sampling. To address these issues, we propose CcGAN-AVAR, an imbalance-aware extension of CcGAN that combines soft/hybrid adaptive vicinity with auxiliary discriminator-guided regularization. The adaptive vicinity constructs a label-dependent local radius according to the available samples around each target condition, and the multi-task discriminator supplies both a regression signal for label consistency and a density-ratio-estimation signal for distribution matching. We further provide a theoretical interpretation characterizing how adaptive vicinal weighting affects the local bias-variance behavior of the discriminator target, how hybrid truncation reduces objective-level cross-condition mixing, and how the density-ratio-based generator penalty approximates a Pearson Chi-square discrepancy up to the estimation error of the density-ratio branch. Extensive experiments on four datasets, including the newly constructed imbalanced RC-49-I, covering resolutions from 64x64 to 256x256 across eleven settings, demonstrate that CcGAN-AVAR obtains strong generation quality and label consistency while preserving the one-step sampling efficiency of GANs, achieving 300x--2000x faster inference than CCDM.

Figures

Figures reproduced from arXiv: 2508.01725 by the authors.

Figure 1
Figure 1. The illustration of CCGM with data imbalance in UTKFace [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparison of traditional fixed-size (hard/soft) vicinities versus [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Training pipeline of the CcGAN-AVAR framework. Novel [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Illustrative example: Building the soft or hybrid adaptive vicinity for condition [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Architectures of the two-layer linear regression branch ( [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Three data imbalance patterns in RC-49-I [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Visual comparison of FID, NIQE, Diversity, and Label Score [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Performance of the strong CcGAN baseline with and without [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Ablation analysis of adaptive vicinity size ( [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 36 canonical work pages

  1. [14]

    CCDM: Continuous conditional diffusion models for image generation,

    X. Ding, Y . Wang, K. Zhang, and Z. J. Wang, “CCDM: Continuous conditional diffusion models for image generation,”arXiv preprint arXiv:2405.03546, 2025

    arXiv 2025

  2. [1]

    PcDGAN: A continuous conditional diverse generative adversarial network for inverse design,

    A. Heyrani Nobari, W. Chen, and F. Ahmed, “PcDGAN: A continuous conditional diverse generative adversarial network for inverse design,” inProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, 2021, pp. 606–616

    work page 2021

  3. [2]

    CcDPM: A continuous conditional diffusion probabilistic model for inverse design,

    Y . Zhao, P. Zhang, G. Sun, Z. Yang, J. Chen, and Y . Wang, “CcDPM: A continuous conditional diffusion probabilistic model for inverse design,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 15, 2024, pp. 17 033–17 041

    work page 2024

  4. [3]

    Diverse 3D auxetic unit cell inverse design with deep learning,

    X. Fang, H.-S. Shen, and H. Wang, “Diverse 3D auxetic unit cell inverse design with deep learning,”Applied Physics Reviews, vol. 10, no. 3, 2023

    work page 2023

  5. [4]

    Data augmentation using continuous conditional generative adversarial networks for regression and its application to improved spectral sensing,

    Y . Zhu, H. Su, P. Xu, Y . Xu, Y . Wang, C.-H. Dong, J. Lu, Z. Le, X. Yang, Q. Xuanet al., “Data augmentation using continuous conditional generative adversarial networks for regression and its application to improved spectral sensing,”Optics Express, vol. 31, no. 23, pp. 37 722– 37 739, 2023

    work page 2023

  6. [5]

    Point cloud generation with continuous conditioning,

    L. T. Triess, A. B¨uhler, D. Peter, F. B. Flohr, and M. Z¨ollner, “Point cloud generation with continuous conditioning,” inInternational Conference on Artificial Intelligence and Statistics, 2022, pp. 4462–4481. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2025 11 TABLE III: Ablation study: Component-wise analysis of CcGAN-A V AR. Dataset Configur...

    work page 2022

  7. [6]

    Predicting the co2 propagation in geological formations from sparsely available well data,

    C. A. Ferreira, M. Stepien, S. Hosseinzadehsadati, T. Kadeethum, and H. M. Nick, “Predicting the co2 propagation in geological formations from sparsely available well data,” in16th International Conference on Greenhouse Gas Control Technologies, 2022

    work page 2022

  8. [7]

    GAN-based framework for unified estimation of process-induced random variation in FinFET,

    T. Park, J. Kwak, H. Ahn, J. Lee, J. Lim, S. Yu, C. Shin, and T. Moon, “GAN-based framework for unified estimation of process-induced random variation in FinFET,”IEEE Access, vol. 10, pp. 130 001–130 023, 2022

    work page 2022

Show all 43 references
  1. [8]

    Scaling up gans for text-to-image synthesis,

    M. Kang, J.-Y . Zhu, R. Zhang, J. Park, E. Shechtman, S. Paris, and T. Park, “Scaling up gans for text-to-image synthesis,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 10 124–10 134

    2023

  2. [9]

    Image de-raining using a conditional generative adversarial network,

    H. Zhang, V . Sindagi, and V . M. Patel, “Image de-raining using a conditional generative adversarial network,”IEEE transactions on circuits and systems for video technology, vol. 30, no. 11, pp. 3943–3956, 2019

    2019

  3. [10]

    Statistics enhancement generative adversarial networks for diverse conditional image synthesis,

    Z. Zuo, A. Li, Z. Wang, L. Zhao, J. Dong, X. Wang, and M. Wang, “Statistics enhancement generative adversarial networks for diverse conditional image synthesis,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 7, pp. 6167–6180, 2024

    2024

  4. [11]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 4195–4205

    2023

  5. [12]

    Adaptive double-branch fusion conditional diffusion model for underwater image restoration,

    Y . Wang, K. He, Q. Qu, X. Du, T. Liu, T. Lei, and A. K. Nandi, “Adaptive double-branch fusion conditional diffusion model for underwater image restoration,”IEEE Transactions on Circuits and Systems for Video Technology, 2025

    2025

  6. [13]

    Image intrinsic components guided conditional diffusion model for low-light image enhancement,

    S. Kang, S. Gao, W. Wu, X. Wang, S. Wang, and G. Qiu, “Image intrinsic components guided conditional diffusion model for low-light image enhancement,”IEEE Transactions on Circuits and Systems for Video Technology, 2024

    2024

  7. [15]

    CcGAN: Continuous conditional generative adversarial networks for image generation,

    X. Ding, Y . Wang, Z. Xu, W. J. Welch, and Z. J. Wang, “CcGAN: Continuous conditional generative adversarial networks for image generation,” inInternational Conference on Learning Representations, 2021

    2021

  8. [16]

    Continuous conditional generative adversarial networks: Novel empirical losses and label input mechanisms,

    ——, “Continuous conditional generative adversarial networks: Novel empirical losses and label input mechanisms,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 7, pp. 8143–8158, 2023

    2023

  9. [17]

    Image generation using continuous conditional generative adver- sarial networks,

    ——, “Image generation using continuous conditional generative adver- sarial networks,” inGenerative Adversarial Learning: Architectures and Applications. Springer, 2022, pp. 87–113

    2022

  10. [18]

    Generative adversarial nets,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial nets,” inAdvances in Neural Information Processing Systems 27, 2014, pp. 2672–2680

    2014

  11. [19]

    Improved distribution matching distillation for fast image synthesis,

    T. Yin, M. Gharbi, T. Park, R. Zhang, E. Shechtman, F. Durand, and W. T. Freeman, “Improved distribution matching distillation for fast image synthesis,” inThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  12. [20]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” inInternational Conference on Learning Representations, 2021

    2021

  13. [21]

    Efficient subsampling of realistic images from GANs conditional on a class or a continuous variable,

    X. Ding, Y . Wang, Z. J. Wang, and W. J. Welch, “Efficient subsampling of realistic images from GANs conditional on a class or a continuous variable,”Neurocomputing, vol. 517, pp. 188–200, 2023

    2023

  14. [22]

    Subsampling generative adversarial networks: Density ratio estimation in feature space with JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2025 12 softplus loss,

    X. Ding, Z. J. Wang, and W. J. Welch, “Subsampling generative adversarial networks: Density ratio estimation in feature space with JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2025 12 softplus loss,”IEEE Transactions on Signal Processing, vol. 68, pp. 1910–1922, 2020

    2025

  15. [23]

    Turning waste into wealth: Leveraging low-quality samples for enhancing continuous conditional generative adversarial networks,

    X. Ding, Y . Wang, and Z. Xu, “Turning waste into wealth: Leveraging low-quality samples for enhancing continuous conditional generative adversarial networks,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 10, 2024, pp. 11 802–11 810

    2024

  16. [24]

    Spectral normal- ization for generative adversarial networks,

    T. Miyato, T. Kataoka, M. Koyama, and Y . Yoshida, “Spectral normal- ization for generative adversarial networks,” inInternational Conference on Learning Representations, 2018

    2018

  17. [25]

    Unsupervised representation learning with deep convolutional generative adversarial networks,

    A. Radford, L. Metz, and S. Chintala, “Unsupervised representation learning with deep convolutional generative adversarial networks,”arXiv preprint arXiv:1511.06434, 2015

    2015 arXiv

  18. [26]

    Self-attention generative adversarial networks,

    H. Zhang, I. Goodfellow, D. Metaxas, and A. Odena, “Self-attention generative adversarial networks,” inProceedings of the 36th International Conference on Machine Learning, 2019, pp. 7354–7363

    2019

  19. [27]

    Large scale GAN training for high fidelity natural image synthesis,

    A. Brock, J. Donahue, and K. Simonyan, “Large scale GAN training for high fidelity natural image synthesis,” inInternational Conference on Learning Representations, 2019

    2019

  20. [28]

    K. P. Murphy,Probabilistic machine learning: an introduction. MIT press, 2022

    2022

  21. [29]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

    2016

  22. [30]

    f-GAN: Training generative neural samplers using variational divergence minimization,

    S. Nowozin, B. Cseke, and R. Tomioka, “f-GAN: Training generative neural samplers using variational divergence minimization,”Advances in neural information processing systems, vol. 29, 2016

    2016

  23. [31]

    Analyzing and improving the training dynamics of diffusion models,

    T. Karras, M. Aittala, J. Lehtinen, J. Hellsten, T. Aila, and S. Laine, “Analyzing and improving the training dynamics of diffusion models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 24 174–24 184

    2024

  24. [32]

    Group normalization,

    Y . Wu and K. He, “Group normalization,”International Journal of Computer Vision, vol. 128, no. 3, pp. 742–755, 2020

    2020

  25. [33]

    Conditional image synthesis with auxiliary classifier gans,

    A. Odena, C. Olah, and J. Shlens, “Conditional image synthesis with auxiliary classifier gans,” inInternational conference on machine learning. PMLR, 2017, pp. 2642–2651

    2017

  26. [34]

    Rebooting ACGAN: Auxiliary classifier GANs with stable training,

    M. Kang, W. Shim, M. Cho, and J. Park, “Rebooting ACGAN: Auxiliary classifier GANs with stable training,”Advances in neural information processing systems, vol. 34, pp. 23 505–23 518, 2021

    2021

  27. [35]

    Conditional GANs with auxiliary discriminative classifier,

    L. Hou, Q. Cao, H. Shen, S. Pan, X. Li, and X. Cheng, “Conditional GANs with auxiliary discriminative classifier,” inProceedings of the 39th International Conference on Machine Learning, 2022, pp. 8888–8902

    2022

  28. [36]

    Mixed preci- sion training,

    P. Micikevicius, S. Narang, J. Alben, G. Diamos, E. Elsen, D. Garcia, B. Ginsburg, M. Houston, O. Kuchaiev, G. Venkateshet al., “Mixed preci- sion training,” inInternational Conference on Learning Representations, 2018

    2018

  29. [37]

    Accelerate: Training and inference at scale made simple, efficient and adaptable

    S. Gugger, L. Debut, T. Wolf, P. Schmid, Z. Mueller, S. Mangrulkar, M. Sun, and B. Bossan, “Accelerate: Training and inference at scale made simple, efficient and adaptable.” https://github.com/huggingface/ accelerate, 2022

    2022

  30. [38]

    Age progression/regression by conditional adversarial autoencoder,

    Z. Zhang, Y . Song, and H. Qi, “Age progression/regression by conditional adversarial autoencoder,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 5810–5818

    2017

  31. [39]

    The Steering Angle dataset @ONLINE,

    S. Chen, “The Steering Angle dataset @ONLINE,” https://github.com/ SullyChen/driving-datasets, 2018

    2018

  32. [40]

    Real-ESRGAN: Training real- world blind super-resolution with pure synthetic data,

    X. Wang, L. Xie, C. Dong, and Y . Shan, “Real-ESRGAN: Training real- world blind super-resolution with pure synthetic data,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 1905–1914

    2021

  33. [41]

    Towards real-world blind face restoration with generative facial prior,

    X. Wang, Y . Li, H. Zhang, and Y . Shan, “Towards real-world blind face restoration with generative facial prior,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 9168– 9178

    2021

  34. [42]

    Differentiable augmentation for data-efficient GAN training,

    S. Zhao, Z. Liu, J. Lin, J.-Y . Zhu, and S. Han, “Differentiable augmentation for data-efficient GAN training,”Advances in Neural Information Processing Systems, vol. 33, pp. 7559–7570, 2020

    2020

  35. [43]

    Making a “completely blind

    A. Mittal, R. Soundararajan, and A. C. Bovik, “Making a “completely blind” image quality analyzer,”IEEE Signal Processing Letters, vol. 20, no. 3, pp. 209–212, 2012. APPENDIX SUPPLEMENTALMATERIAL The source code and implementation details will be made publicly available at: ht...

    2012

Pith tools

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