Pith. sign in

REVIEW 4 major objections 5 minor 37 references

Deep Generative Methods and Tire Architecture Design

T0 review · 4 major / 5 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 This paper benchmarks five deep generative model families on industrial tire architecture design and concludes that diffusion models are the strongest overall, while a masking-trained VAE beats the more complex multimodal VAE on…

desk verdict A solid empirical benchmark with one overstated generalization claim; the DDPM-vs-MDM OOD gap rests on a test-selected guidance scale. read the letter →

arxiv 2507.11639 v2 pith:B6OFITOS submitted 2025-07-15 cs.LG

Fouad Oubari , Raphael Meunier , Rodrigue Décatoire , Mathilde Mougeot This is my paper · ORCID
classification cs.LG
keywords deepgenerativemodelsdiffusionvariationalautoencodersadversarialnetworkstirearchitecturedesigncategoricalinpaintingmultinomialconditionalgeneration
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

Industrial design teams need to know which deep generative model family to use when generating multi-component products. This paper tries to answer that question by benchmarking five representative models—variational autoencoder (VAE), generative adversarial network (GAN), multimodal VAE (MMVAE+), denoising diffusion probabilistic model (DDPM), and multinomial diffusion model (MDM)—under matched parameter budgets on a dataset of 1,130 tire architectures, each encoded as six binary 64×64 component masks. The central finding is that diffusion models are the strongest overall: MDM leads on in-distribution fidelity, DDPM generalizes better to out-of-distribution dimensional constraints, and a masking-trained VAE outperforms the multimodal MMVAE+ on nearly all component-conditioned metrics. The paper also introduces categorical inpainting, a mask-aware reverse diffusion step that lets discrete diffusion models condition on observed components without additional training. If the ranking holds, it gives practitioners a concrete model-selection heuristic and a training-free way to make discrete diffusion conditional.

What carries the argument

The load-bearing mechanism is categorical inpainting, a mask-aware reverse diffusion step for multinomial diffusion. Given the learned reverse kernel $p(c_{t-1}|c_t) = \mathrm{Cat}(c_{t-1}|\theta)$ from the multinomial diffusion model, the paper multiplies the per-pixel probability vector by a binary mask $r$ that is one-hot for observed labels and uniform elsewhere; Proposition 1 shows the renormalized kernel is again categorical with $\tilde{\theta}_k = r_k \theta_k / \sum_j r_j \theta_j$. This is what lets an already-trained MDM generate conditionally on a known component or a dimensional specification without retraining. The other machinery is the controlled benchmark itself: five models with comparable parameter counts, a shared training protocol, and geometry-aware metrics (IoU, center of mass, region-connectivity error, FID, dimension error) that map to industrial requirements.

What would settle it

Retrain the same five models on a version of the dataset that encodes continuous material thickness (or full 3D geometry) instead of binary 64×64 masks, under the same parameter budgets and metric suite; if the diffusion models no longer dominate or the masking-trained VAE no longer beats MMVAE+, the paper's ranking is an artifact of the binary-mask proxy rather than a property of multi-component design generation.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is a performance hierarchy that is consistent across three industrial scenarios. In unconditional generation, the multinomial diffusion model MDM and the denoising diffusion model DDPM produce clean, spatially coherent architectures, while the GAN collapses and the VAEs produce duplication or blur; diffusion models win on the center-of-mass, region-connectivity, and FID metrics. In component-conditioned generation, the masking-trained VAE beats MMVAE+ on nearly all metrics, a result the paper attributes to its single six-channel input, early feature sharing, and progressive masking curriculum. In dimension-constrained generation, MDM has lower FID and dimension error in-distribution, while DDPM produces more plausible architectures for out-of-distribution specifications. Methodologically, the paper contributes categorical inpainting: reweighting each reverse diffusion step by a binary mask that freezes observed labels, with Proposition 1 showing the masked reverse kernel remains a categorical distribution.

Load-bearing premise

The ranking rests on the assumption that 1,130 tire architectures, encoded as six binary 64×64 component masks, faithfully represent the tire design space; if continuous material thickness, 3D shape, or manufacturing constraints are what actually matter, the model ranking may not carry over to real design work.

Editorial extensions

If this is right

  • Design teams working on multi-component products should treat diffusion models as the default generative family for unconditional and conditional design generation, rather than VAE or GAN variants.
  • Within the diffusion family, MDM is the better choice when the priority is high-fidelity reproduction of known designs, while DDPM is the better choice when the goal is exploring dimensional specifications outside the training envelope.
  • A single-channel masking curriculum turns a plain VAE into a competitive conditional generator, beating a more elaborate multimodal VAE on most component-conditioned metrics at lower complexity.
  • Categorical inpainting gives discrete diffusion models conditional-generation capability for free, so an existing MDM can be reused for new conditioning tasks without retraining.
  • The geometry-aware metric suite is dataset-agnostic and transfers to other multi-component design problems, giving later benchmarks a shared evaluation protocol.

Reading between the lines

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

  • Because the paper identifies separate encoders and the absence of a masking curriculum as the reasons MMVAE+ lags, a natural extension is to add progressive masking to MMVAE+'s training; if the gap closes, the paper's ranking of conditional models would change.
  • The MDM/DDPM split on in-distribution versus out-of-distribution suggests a general trade-off for industrial generative modeling: categorical discrete models overfit the training classes, while continuous Gaussian diffusion has more room to extrapolate, a hypothesis testable on other structured design datasets.
  • Categorical inpainting is not tire-specific: any one-hot categorical generation task, such as semantic map synthesis or discrete material assignment, can use the same masked reverse step for training-free conditioning.
  • The OOD comparison drops FID, so the DDPM OOD advantage rests only on connectivity and dimension-error metrics; a perceptual metric that works outside the training distribution could alter the apparent winner.
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. This paper reports a controlled comparison of five deep generative model families (VAE, GAN, MMVAE+, DDPM, and MDM) on a proprietary dataset of 1,130 tire architectures represented as six binary 64x64 component masks. The authors evaluate unconditional generation, component-conditioned generation, and dimension-constrained generation using geometry-aware metrics (IoU, CoM, RCE, FID, DimErr). For MDM they introduce a masked categorical reverse-diffusion step ('categorical inpainting') that preserves observed labels without retraining, and they provide a proposition and proof that the masked kernel remains categorical. The main empirical findings are that diffusion models outperform VAE/GAN baselines on most metrics; that a progressively masked VAE outperforms MMVAE+ on component-conditioned metrics; and that within the diffusion family MDM leads in-distribution while DDPM generalizes better to out-of-distribution dimensional constraints, under classifier-free guidance with a single guidance weight.

Significance. If the findings hold, the paper offers a useful reference point for practitioners choosing generative models for multi-component industrial design, and the proposed categorical inpainting for MDM is a simple and sound extension. Strengths include the controlled parameter budgets, three random seeds per model, multiple complementary metrics, and a mathematically verified masked reverse-step construction. The methodological contribution of categorical inpainting is clearly stated and its proof is straightforward. The main limitations are the small proprietary dataset, the absence of a validation split for selecting the guidance scale used in the out-of-distribution claim, and the mixed OOD evidence, which together mean that the strongest generalization claim in the abstract is not yet fully supported.

major comments (4)
  1. [Appendix G and Section IV-D] The out-of-distribution (OOD) comparison between DDPM and MDM is not robust because the classifier-free guidance scale used in the OOD evaluation was selected on the in-distribution test set. Appendix G states that w was varied over {1,...,10} and w=4 was chosen empirically to minimize FID while maximizing DimErr 'in the in-distribution setting', with no validation split and no OOD sensitivity analysis. Since the same w=4 is carried into the OOD runs reported in Table VI, the DDPM-vs-MDM OOD ranking may be an artifact of this choice. Please tune w on a held-out validation split and report that protocol, or provide an OOD sweep over w showing that the ranking is stable.
  2. [Section IV-D and Table VI] The OOD quantitative evidence does not uniformly support the abstract claim that 'DDPM generalises better to out-of-distribution dimensional constraints.' In the OOD split, DDPM is better on RCE (15.485 vs 36.978) and DimErr (9.528 vs 11.795), but MDM is substantially better on CoM (3.986 vs 7.126). No aggregate test or statistical comparison is reported, and the RCE standard deviations overlap substantially (15.485 +/- 11.175 vs 36.978 +/- 13.138). I recommend either softening the abstract and Section IV-D to a tradeoff statement, or adding a principled multi-metric comparison (e.g., dominance analysis or paired significance test) that actually supports the generalization claim.
  3. [Section IV and Table VI] The conclusion that diffusion models achieve the strongest overall performance is only partially supported because dimension-constrained generation is evaluated exclusively with DDPM and MDM. Section IV states that this restriction follows from the 'clear superiority of diffusion-based models observed in unconstrained generation,' but the component-conditioned experiments show that a masking-trained VAE can be competitive with (and even surpass) diffusion models on RCE. Excluding VAE and MMVAE+ from the dimension-constrained scenario leaves the overall-ranking claim incomplete. Please include at least the masking-trained VAE under dimension constraints, or reframe the conclusion as an unconditional/conditional ranking plus a diffusion-only dimension-constrained comparison.
  4. [Section III-B and Tables II-VI] The empirical rankings rest on a single proprietary dataset of 1,130 samples with an 80/20 split and three random seeds. Table VI shows that between-model differences on several key metrics are within or close to the reported standard deviations (notably OOD RCE). Please add an explicit limitations paragraph discussing the small dataset size and potential transferability issues, and, if possible, report repeated-split or per-seed results so that the stability of the rankings can be assessed.
minor comments (5)
  1. [Section III-B and Appendix G] The MDM preprocessing description is inconsistent: Section III-B calls the input a '6-channel one-hot tensor' with six channels, while Section III-A defines K = M+1 = 7 classes and Appendix G says the six modalities are collapsed into a single categorical map with K+1 values. Please reconcile these statements.
  2. [Table I] The text says the models have comparable parameter counts, but MMVAE+ has substantially higher GFLOPs (18.77) than DDPM (11.96) and MDM (9.87). The 'identical parameter budgets' claim should be accompanied by a discussion of the FLOPs asymmetry, since both parameters and compute can affect performance.
  3. [Section IV-D] The sentence 'we evaluate models the four metrics' appears to contain a typo; it should read 'we evaluate the models on the four metrics.'
  4. [Appendix G] The DDPM training objective is described as 'the pred-v objective,' but the main text Section III-C defines DDPM as predicting the noise epsilon. Please clarify whether v-prediction or epsilon-prediction is used, and align the notation.
  5. [Figure 6] The caption says 'Points represent individual training runs with different random seeds,' while the appendix says every entry is the mean +/- standard deviation over three runs. Please clarify how the points in Figure 6 relate to the values in Tables II-V.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the benchmark comparisons are self-contained empirical results, Proposition 1 is a straightforward proven reweighting of an external kernel, and the few self-citations are non-load-bearing.

full rationale

This paper is a controlled empirical benchmark whose derivation chain is self-contained. The only formal derivation, Proposition 1 (masked categorical reverse step), reweights the multinomial reverse kernel from Hoogeboom et al.'s Eq. (14) by a non-negative mask and renormalizes; the proof in Appendix D is a direct normalization argument, so the result follows from the cited external kernel without importing the paper's own conclusion. The central claims - diffusion models outperform VAE/GAN families, and the masking-trained VAE outperforms MMVAE+ on component-conditioned metrics - rest on metrics (FID, CoM, RCE, IoU, DimErr) computed in-paper on held-out test data, with mean plus/minus standard deviation over three independent seeds; no fitted parameter is renamed as a prediction. The only self-citations ([16] Binded-VAE, [17] Meta-VAE) appear as contextual related-work examples and carry no load-bearing quantitative claim. Two methodological caveats should be recorded but do not rise to circularity. First, Appendix G states that the guidance weight w was varied over {1,...,10} and that w=4 was chosen empirically to minimize FID while maximizing DimErr in the in-distribution setting, with no validation split described, and that same w=4 is then carried into the OOD evaluation; the abstract's claim that 'DDPM generalises better to out-of-distribution dimensional constraints' is therefore sensitive to that test-selected hyperparameter and is in fact only partially supported by Table VI, where DDPM wins OOD RCE and DimErr while MDM wins OOD CoM (3.986 vs 7.126). This is an overstatement and a hyperparameter-selection concern, not a by-construction reduction. Second, MDM's categorical output forbids pixel overlap by construction, which is disclosed at the IoU metric definition and structures a portion of its spatial-coherence advantage; however, MDM's leading CoM, RCE, and FID scores are not forced by that property, and the paper explicitly excludes MDM from IoU rather than claiming a win on it. Accordingly, no circular step can be exhibited, and the appropriate finding is essentially no circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central comparison rests on a small number of hand-selected hyperparameters and on unvalidated domain assumptions about the representativeness of the proprietary dataset and the meaningfulness of ImageNet-based FID for tire cross-sections. No new physical entities are introduced.

free parameters (4)
  • VAE regularization weight beta = 2.5
    Selected by grid search over {0.01, 0.1, 1, 2.5, 5} in Appendix G; the masking-trained VAE's conditional-generation performance depends on this value.
  • VAE latent dimension dz = 48
    Selected from {32, 48, 64} in Appendix G; controls the information bottleneck and affects reconstruction quality.
  • Progressive masking ceiling = 0.7 with warmup over 20 epochs
    Hand-chosen curriculum in Appendix B; the claim that masking-trained VAE outperforms MMVAE+ under conditioning rests on this schedule.
  • Classifier-free guidance scale w = 4
    Chosen by scanning w in {1,...,10} and optimizing FID and DimErr on the in-distribution test set (Appendix G); directly influences dimension-constrained generation results.
assumptions (5)
  • domain assumption The 64x64 binary-mask cross-section representation with six components is a sufficient description of tire architecture for generative design comparisons.
    Section III-A and III-B. The entire benchmark operates on this representation; if real tires require 3D or continuous material information, rankings may not transfer.
  • domain assumption The 1,130 proprietary architectures are representative of the tire design space and the 80/20 split yields stable evaluation statistics.
    Section III-B. Small sample size and proprietary sourcing limit generalizability; no external dataset is used for validation.
  • domain assumption FID computed with an ImageNet-pretrained Inception network is a meaningful perceptual-fidelity measure for 64x64 grayscale tire cross-sections.
    Section IV-A.b. The validity of natural-image features for industrial cross-sections is assumed, not demonstrated.
  • standard math 1-Wasserstein distances between empirical metric distributions are reliably estimated from the generated sample sizes.
    Section IV-A.b and Tables II-VI; W1 estimates are reported with standard deviations over three runs, but conditioned-generation sample sizes are small.
  • domain assumption The convex hull of the training-set (width, height) distribution defines the in-distribution versus out-of-distribution boundary.
    Section III-B and Figure 2; conclusions about OOD generalization depend on this partition.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Generative Methods and Tire Architecture Design." pith.science (2026). https://pith.science/paper/B6OFITOS

@misc{pith2026250711639,
  author       = {Pith},
  title        = {Pith review of: Deep Generative Methods and Tire Architecture Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B6OFITOS}},
  note         = {Machine review of arXiv:2507.11639}
}
read the original abstract

As deep generative models proliferate across the AI landscape, industrial practitioners still face critical yet unanswered questions about which deep generative models best suit complex manufacturing design tasks. This work addresses this question through a complete study of five representative models (Variational Autoencoder, Generative Adversarial Network, multimodal Variational Autoencoder, Denoising Diffusion Probabilistic Model, and Multinomial Diffusion Model) on industrial tire architecture generation. Our evaluation spans three key industrial scenarios: (i) unconditional generation of complete multi-component designs, (ii) component-conditioned generation (reconstructing architectures from partial observations), and (iii) dimension-constrained generation (creating designs that satisfy specific dimensional requirements). To enable discrete diffusion models to handle conditional scenarios, we introduce categorical inpainting, a mask-aware reverse diffusion process that preserves known labels without requiring additional training. Our evaluation employs geometry-aware metrics specifically calibrated for industrial requirements, quantifying spatial coherence, component interaction, structural connectivity, and perceptual fidelity. Our findings reveal that diffusion models achieve the strongest overall performance; a masking-trained VAE nonetheless outperforms the multimodal variant MMVAE\textsuperscript{+} on nearly all component-conditioned metrics, and within the diffusion family MDM leads in-distribution whereas DDPM generalises better to out-of-distribution dimensional constraints.

Figures

Figures reproduced from arXiv: 2507.11639 by the authors.

Figure 1
Figure 1. Tire architecture representation used in this work. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Distribution of tire dimensional attributes (height and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Qualitative results of unconditional generations for the 5 benchmarked DGMs. For each model, the first five columns [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Qualitative results of component-constrained conditional generation. The figure shows two examples (top and bottom [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: In-distribution comparison of dimension-constrained [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Performance comparison of generative models across four metrics (IoU, CoM, RCE, FID) for unconstrained generation [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: OOD sample comparison of dimension-constrained generation. The left plot shows the distribution hull from Figure [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Quantitative performance of diffusion models on [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 22 canonical work pages

  1. [1]

    Auto-encoding variational bayes,

    D. P. Kingma, M. Welling et al. , “Auto-encoding variational bayes,” 2013

    2013

  2. [2]

    Generative adversarial nets,

    I. J. 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 , vol. 27, 2014

    2014

  3. [3]

    Denoising diffusion probabilistic models,

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

    2020

  4. [4]

    Shapenet: An information- rich 3d model repository,

    A. X. Chang, T. Funkhouser, L. Guibas, P. Hanrahan, Q. Huang, Z. Li, S. Savarese, M. Savva, S. Song, H. Su et al., “Shapenet: An information- rich 3d model repository,” arXiv preprint arXiv:1512.03012 , 2015

    arXiv 2015

  5. [5]

    Mmvae+: Enhancing the generative quality of multimodal vaes without compromises,

    E. Palumbo, I. Daunhawer, and J. E. V ogt, “Mmvae+: Enhancing the generative quality of multimodal vaes without compromises,” in The Eleventh International Conference on Learning Representations . OpenReview, 2023

    2023

  6. [6]

    Argmax flows and multinomial diffusion: Learning categorical distributions,

    E. Hoogeboom, D. Nielsen, P. Jaini, P. Forr ´e, and M. Welling, “Argmax flows and multinomial diffusion: Learning categorical distributions,” Advances in neural information processing systems, vol. 34, pp. 12 454– 12 465, 2021

    work page 2021

  7. [7]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,” Advances in neural information processing systems , vol. 30, 2017

    2017

  8. [8]

    Towards high-fidelity cfd on the cloud for the automotive and motorsport sectors,

    N. Ashton, S. Sachs, L. Foti, and S. Eberhardt, “Towards high-fidelity cfd on the cloud for the automotive and motorsport sectors,” SAE Technical Paper, Tech. Rep., 2020

    work page 2020

Show all 37 references
  1. [9]

    Modeling and validation of a passenger car tire using finite element analysis,

    H. Fathi, Z. El-Sayegh, J. Ren, and M. El-Gindy, “Modeling and validation of a passenger car tire using finite element analysis,” Vehicles, vol. 6, no. 1, pp. 384–402, 2024

    2024

  2. [10]

    Comparison of optimization algorithms for aerodynamic shape design,

    S. Obayashi and T. Tsukahara, “Comparison of optimization algorithms for aerodynamic shape design,” AIAA journal, vol. 35, no. 8, pp. 1413– 1415, 1997

    1997

  3. [11]

    Efficient global optimiza- tion of expensive black-box functions,

    D. R. Jones, M. Schonlau, and W. J. Welch, “Efficient global optimiza- tion of expensive black-box functions,” Journal of Global optimization , vol. 13, pp. 455–492, 1998

    1998

  4. [12]

    A. E. Eiben and J. E. Smith, Introduction to evolutionary computing . Springer, 2015

    2015

  5. [13]

    Devel- opment of a conditional generative adversarial network for airfoil shape optimization,

    G. Achour, W. J. Sung, O. J. Pinon-Fischer, and D. N. Mavris, “Devel- opment of a conditional generative adversarial network for airfoil shape optimization,” in AIAA Scitech 2020 Forum , 2020, p. 2261

    2020

  6. [14]

    Generating various airfoils with required lift coefficients by combining naca and joukowski airfoils using conditional variational autoencoders,

    K. Yonekura, K. Wada, and K. Suzuki, “Generating various airfoils with required lift coefficients by combining naca and joukowski airfoils using conditional variational autoencoders,” Engineering Applications of Artificial Intelligence, vol. 108, p. 104560, 2022

    2022

  7. [15]

    Variational autoencoder- based topological optimization of an anechoic coating: An efficient- and neural network-based design,

    Y . Sun, Z. Li, J. Chen, X. Zhao, and M. Tao, “Variational autoencoder- based topological optimization of an anechoic coating: An efficient- and neural network-based design,” Materials Today Communications , vol. 32, p. 103901, 2022

    2022

  8. [16]

    A binded vae for inorganic material generation,

    F. Oubari, A. De Mathelin, R. D ´ecatoire, and M. Mougeot, “A binded vae for inorganic material generation,” arXiv preprint arXiv:2112.09570, 2021

    2021 arXiv

  9. [17]

    A meta- vae for multi-component industrial systems generation,

    F. Oubari, R. Meunier, R. D ´ecatoire, and M. Mougeot, “A meta- vae for multi-component industrial systems generation,” in Intelligent Computing, K. Arai, Ed. Cham: Springer Nature Switzerland, 2024, pp. 234–251

    2024

  10. [18]

    A survey of multimodal deep generative models,

    M. Suzuki and Y . Matsuo, “A survey of multimodal deep generative models,” Advanced Robotics, vol. 36, no. 5-6, pp. 261–278, 2022

    2022

  11. [19]

    Multimodal generative models for scalable weakly-supervised learning,

    M. Wu and N. Goodman, “Multimodal generative models for scalable weakly-supervised learning,” Advances in neural information processing systems, vol. 31, 2018

    2018

  12. [20]

    Variational mixture-of-experts autoen- coders for multi-modal deep generative models,

    Y . Shi, B. Paige, P. Torr et al., “Variational mixture-of-experts autoen- coders for multi-modal deep generative models,” Advances in neural information processing systems , vol. 32, 2019

    2019

  13. [21]

    Multimodal generative learning utilizing jensen-shannon-divergence,

    T. Sutter, I. Daunhawer, and J. V ogt, “Multimodal generative learning utilizing jensen-shannon-divergence,” Advances in neural information processing systems, vol. 33, pp. 6100–6110, 2020

    2020

  14. [22]

    Generalized multimodal elbo,

    T. M. Sutter, I. Daunhawer, and J. E. V ogt, “Generalized multimodal elbo,” arXiv preprint arXiv:2105.02470 , 2021

    2021 arXiv

  15. [23]

    Material microstructure design using vae-regression with a multimodal prior,

    A. Sardeshmukh, S. Reddy, B. Gautham, and P. Bhattacharyya, “Material microstructure design using vae-regression with a multimodal prior,” in Pacific-Asia Conference on Knowledge Discovery and Data Mining . Springer, 2024, pp. 29–41

    2024

  16. [24]

    Psp-gen: Stochastic inversion of the process–structure–property chain in materials design through deep, generative probabilistic modeling,

    Y . Zang and P.-S. Koutsourelakis, “Psp-gen: Stochastic inversion of the process–structure–property chain in materials design through deep, generative probabilistic modeling,” Acta Materialia, vol. 284, p. 120600, 2025

    2025

  17. [25]

    Score-based generative modeling through stochastic differ- ential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differ- ential equations,” arXiv preprint arXiv:2011.13456 , 2020

    2011 arXiv

  18. [26]

    Diffusion models beat gans on topology optimization,

    F. Maz ´e and F. Ahmed, “Diffusion models beat gans on topology optimization,” in Proceedings of the AAAI conference on artificial intelligence, vol. 37, no. 8, 2023, pp. 9108–9116

    2023

  19. [27]

    A data-driven framework for designing microstructure of multifunctional composites with deep- learned diffusion-based generative models,

    K.-H. Lee, H. J. Lim, and G. J. Yun, “A data-driven framework for designing microstructure of multifunctional composites with deep- learned diffusion-based generative models,” Engineering Applications of Artificial Intelligence, vol. 129, p. 107590, 2024

    2024

  20. [28]

    Benchmarking study of deep generative models for inverse polymer design,

    T. Yue, L. Tao, V . Varshney, and Y . Li, “Benchmarking study of deep generative models for inverse polymer design,” Digital Discovery, 2025

    2025

  21. [29]

    Generative models struggle with kirigami metamaterials,

    G. Felsch and V . Slesarenko, “Generative models struggle with kirigami metamaterials,” Scientific Reports, vol. 14, no. 1, p. 19397, 2024

    2024

  22. [30]

    Dismai-bench: benchmarking and designing generative models using disordered materials and interfaces,

    A. X. B. Yong, T. Su, and E. Ertekin, “Dismai-bench: benchmarking and designing generative models using disordered materials and interfaces,” Digital Discovery, vol. 3, no. 9, pp. 1889–1909, 2024

    1909

  23. [31]

    Vibration-based anomaly detection in industrial machines: A comparison of autoencoders and latent spaces,

    L. Radicioni, F. M. Bono, and S. Cinquemani, “Vibration-based anomaly detection in industrial machines: A comparison of autoencoders and latent spaces,” Machines, vol. 13, no. 2, p. 139, 2025

    2025

  24. [32]

    Deep unsupervised learning using nonequilibrium thermodynamics,

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermodynamics,” in International conference on machine learning . pmlr, 2015, pp. 2256– 2265

    2015

  25. [33]

    Classifier-free diffusion guidance,

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,” arXiv preprint arXiv:2207.12598, 2022

    2022 arXiv

  26. [34]

    Generalized intersection over union: A metric and a loss for bounding box regression,

    H. Rezatofighi, N. Tsoi, J. Gwak, A. Sadeghian, I. Reid, and S. Savarese, “Generalized intersection over union: A metric and a loss for bounding box regression,” in Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition , 2019, pp. 658–666

    2019

  27. [35]

    Resampled priors for variational autoencoders,

    M. Bauer and A. Mnih, “Resampled priors for variational autoencoders,” in The 22nd International Conference on Artificial Intelligence and Statistics. PMLR, 2019, pp. 66–75

    2019

  28. [36]

    beta-vae: Learning basic visual concepts with a constrained variational framework,

    I. Higgins, L. Matthey, A. Pal, C. Burgess, X. Glorot, M. Botvinick, S. Mohamed, and A. Lerchner, “beta-vae: Learning basic visual concepts with a constrained variational framework,” in International conference on learning representations , 2017. APPENDIX A. Detailed Results T...

    2017

  29. [37]

    Choose k to mirror the encoder depth from Table VIII

    s0×s0 s0 = 8 Reshape View → (n, nf0, s0, s0) 8 ×8 – Repeat k times (reverse order): ResNet Block same Upsample ×2 2 s nearest or transposed conv Tail ResNet Block 64×64 nf→nf Output Conv Conv2d (nf → cout, 3×3) 64 ×64 cout = 1(MMV AE+) or 6 (others) TABLE IX: Generic decoder/g...

Pith tools

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