Pith. sign in

REVIEW 4 major objections 6 minor 49 references

Masked Conditioning for Deep Generative Models

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that masking each condition during training lets a single generative model handle arbitrary sparse, mixed-type conditioning at inference, and demonstrates the idea on small engineering datasets in both VAE and latent…

desk verdict A practical, reproducible engineering paper on masking conditions for sparse mixed-type data, but the core idea is conditioning dropout, and the refinement claim overreaches the paper's own numbers. read the letter →

arxiv 2505.16725 v1 pith:6KODO5U2 submitted 2025-05-22 cs.LG cs.CV

Phillip Mueller , Jannik Wiese , Sebastian Mueller , Lars Mikelsons This is my paper · ORCID
classification cs.LGcs.CV
keywords maskedconditioningsparsemixed-typedatalatentdiffusionmodelsvariationalautoencoderssparsityschedulingengineeringdesigndeepgenerative
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

The paper introduces masked conditioning, a training-time procedure that randomly masks each categorical and numerical condition before embedding, so that at inference the same generative model can condition on any subset of available inputs. This directly targets engineering datasets, which are typically small, sparsely labeled, and contain both categorical and numerical features. The authors show the method works in a lightweight VAE and in a latent diffusion model, and that sparsity schedules act as a curriculum to make the model robust across sparsity levels. If correct, a single model replaces the need for separate models per conditioning subset and lowers the data barrier for generative design.

What carries the argument

The central machinery is the masking step combined with the sparsity scheduler. Each condition is independently replaced by a placeholder with probability $p_t$ during training — a reserved extra token for categoricals, the value $-1$ for numericals — and the scheduler $f:[0,T]\to[0,1]$ (constant, step, linear, or exponential, increasing or decreasing) sets $p_t$ per gradient update. The embedded condition vector $\mathbf{e}_y=[\mathbf{e}_{\text{cat}},\mathbf{e}_{\text{num}}]$ is concatenated either with the latent code in the VAE decoder or, after a learned projection and batch normalization, with the feature maps before each resolution level of the U-Net in the LDM. This mechanism makes inference-time sparsity a point on the training distribution, so the model is trained to handle every subset without retraining.

What would settle it

Train the same architecture on a dataset where missingness is structured or informative — for example, whole condition groups missing together, or where a numerical condition can be negative or zero so the $-1$ sentinel collides with a real value — and compare generation accuracy against the random-masking model; significant degradation would show that the masking simulation does not cover the missingness patterns the method claims to handle.

Watch

Extended reading notes

Core claim

The authors claim that randomly masking conditions during training with a chosen sparsity schedule lets a generative model learn to condition on an arbitrary subset of categorical and numerical inputs at inference time. They implement this by embedding each condition — learnable lookup for categoricals, a linear projection for numericals — concatenating the embeddings, and injecting the result into either a VAE decoder or, in a latent diffusion model, into the U-Net ResBlocks at each resolution. On point-cloud data (GeoBIKED and a proprietary vehicle dataset) and image data (GeoBIKED images and DVM-Car), they report that reconstruction degrades gracefully with sparsity, that roughly 500 training samples suffice for competitive accuracy with the VAE, and that a small masked-conditioned LDM can act as a domain prior whose output is then refined by pretrained models like SDXL and FLUX to add photorealism while retaining controllability.

Load-bearing premise

The load-bearing premise is that independently masking each condition with probability $p_t$ during training faithfully simulates the arbitrary sparse or missing conditions the model will see at inference; the paper only tests random per-condition masking, and the numerical placeholder $-1$ further assumes every numerical condition is strictly positive.

Editorial extensions

If this is right

  • A single masked-conditioned model replaces a family of models trained for each conditioning subset, since inference-time sparsity is a trained input.
  • Small engineering datasets around 500 samples suffice for competitive point-cloud generation accuracy, lowering the data barrier for generative design tools.
  • Sparsity schedules act as a tunable curriculum: increasing sparsity can help very small datasets, while constant low sparsity is often best when full conditions are available at inference.
  • Domain-specific LDMs trained cheaply can be paired with large pretrained image generators, giving photorealism without losing the controllability of the small conditioning scheme.
  • Injecting conditions via concatenation rather than cross-attention reduces computational cost while still supporting mixed-type sparse conditions.

Reading between the lines

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

  • If the random-masking simulation transfers, the same recipe could be applied to text or image conditions by masking tokens or patches, giving a unified way to handle heterogeneous missing conditions.
  • The $-1$ numerical sentinel is an obstacle to generalization: replacing it with a learned mask embedding for numericals would remove the strict-positivity assumption and likely improve robustness on real design variables that can be negative or zero.
  • The finding that training sparsity around $0.5$ can improve fully-conditioned inference suggests masking acts as a regularizer; an ablation isolating this regularization effect from the sparsity-simulation effect would sharpen the design rule.
  • The two-stage refinement result hints that small domain-specific priors plus frozen large models could be a general pattern for scarce-data engineering domains, though the paper only shows it for images.
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 / 6 minor

Summary. The paper introduces masked conditioning for deep generative models: during training, each categorical or numerical conditioning variable is randomly masked with probability p_t, so that at inference the model can generate from arbitrary subsets of mixed-type conditions. The authors integrate the scheme into a VAE and a latent diffusion model, test constant and scheduled sparsity levels on GeoBIKED, a proprietary vehicle point-cloud dataset, and DVM-Car image data, and propose a two-stage pipeline in which a small domain-specific mcLDM produces an image that is refined by SDXL or FLUX. The abstract claims three headline results: handling sparse mixed-type conditioning, data efficiency on small datasets, and refinement with pretrained foundation models that improves generation quality while retaining controllability.

Significance. If fully substantiated, the method would be practically useful for engineering-design settings where annotations are sparse and heterogeneous, and the public code/data release is a positive feature. The VAE experiments are systematic in dataset size and training sparsity, and the architecture choice of conditioning only the decoder is clearly described. However, the paper currently lacks the comparison and statistics needed to establish the core claims: there is no baseline without masking, no error bars or repeated-seed variability, and the refinement claim is quantitatively contradicted by the paper's own Table 4. The novelty relative to conditioning dropout also needs explicit discussion. With those issues addressed, the contribution would be a solid empirical study rather than a fully supported new method.

major comments (4)
  1. [Abstract and Section 4.2, Table 4] The headline claim that coupling small models with large pretrained models improves generation quality while retaining controllability is contradicted by the reported numbers. On DVM-Car, SDXL refinement worsens every reported metric relative to mcLDM alone (MSE 0.0730 vs 0.0678, CLIP-similarity 0.8709 vs 0.9010, SSIM 0.5703 vs 0.5862, LPIPS 0.4067 vs 0.3642), and on GeoBiked SDXL improves MSE only marginally while degrading the other three metrics; FLUX degrades all metrics on both datasets. The text itself acknowledges that using only mcLDM yields better alignment with ground truth. Since 'improve generation quality' and 'retaining controllability' are central to the contribution, this claim must be removed or re-evidenced with metrics that actually measure the claimed photorealism benefit (e.g., FID or a user study) rather than only qualitative samples.
  2. [Section 4.1, Table 2 and Figure 2] The data-efficiency claim that approximately 500 samples is the threshold beyond which further data brings minimal improvement is not supported by the reported table. Table 2 reports dataset sizes 10, 100, 1000, 2000, and 3000, with mean MSEs 0.2398, 0.1217, 0.0998, 0.0914, and 0.0939; there is no 500-sample row, and the mean still improves from 1000 to 2000 samples. The threshold is therefore asserted rather than demonstrated, and the 'minimal improvements' claim needs either the missing experiments or a more modest statement.
  3. [Sections 4.1 and 4.2, Tables 2-4] All quantitative claims are reported without error bars, confidence intervals, or repeated-seed statistics, and no comparison is made to a baseline model trained without masking or to a standard conditional model with full conditioning. Without such a baseline, the paper cannot establish that masking is what enables sparse-condition generation or that the method is data-efficient; the observed behavior could in principle be reproduced by simpler conditioning dropout or by an unconditioned model. Adding at least one unconditioned/fully-conditioned comparison and repeated-seed variability is necessary for the main empirical conclusions.
  4. [Section 3.2] The masking scheme assumes that random independent per-condition masking during training is sufficient for arbitrary sparsity at inference, and it represents missing numerical values with -1 under the stated assumption that all numerical variables are strictly positive. The experiments only test random per-condition masking, not structured or correlated missingness, and the sentinel choice fails for real design variables that can be zero or negative. The paper should either test structured missingness or explicitly scope the claim to the random-missingness, strictly-positive setting used here.
minor comments (6)
  1. [Abstract] There is a typo in the abstract: 'Additionally.' should be 'Additionally,' or should be joined to the following sentence.
  2. [Section 3.1, Eq. (3)] The notation d_num,i is inconsistent with the text's d_num; this should be unified.
  3. [Section 4.1, Dataset Size paragraph] The text refers to 'Table 4' when summarizing dataset-size results, but the relevant table is Table 2; the cross-reference needs correcting.
  4. [Figure 2 caption] The caption says 'BIKED dataset' while the text and the paper's own dataset name are GeoBIKED; this should be made consistent.
  5. [Section 3.2 and Related Work] The novelty of the masking scheme relative to standard conditioning dropout should be discussed explicitly, since the mechanism of randomly dropping conditioning inputs is closely related to existing dropout practice.
  6. [Section 4.2, Image Refinement] The phrase 'significant enhancement in photorealism and perceived image quality' is supported only by qualitative samples; no quantitative perceptual-quality metric is reported for the refinement outputs.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's claims are empirical demonstrations, and its self-citations are not load-bearing.

full rationale

The paper does not derive a prediction from fitted inputs. Its contribution is a training-time masking procedure (Section 3.2), an embedding construction (Section 3.1), and two architecture integrations (Sections 3.3 and 3.4); the claims are supported by held-out measurements on GeoBIKED, a proprietary vehicle dataset, and DVM-Car (Tables 2-4). The masking distribution used at inference is the same family used at training, which is a design assumption rather than a circular reduction: the models still must generalize to held-out samples and to sparsity levels not explicitly trained for, and this is what the experiments test. Self-citations (GeoBIKED [24], InsertDiffusion [25], and [23]) appear as dataset/context references and do not carry the argument. The refinement claim is weakly supported because Table 4 shows SDXL and FLUX degrade most alignment metrics relative to mcLDM, and the paper itself concedes "only using mcLDM results in better alignment with ground truth images"; that is an internal evidence-consistency problem, not a circularity problem. No step of the paper reduces by construction to its own inputs, so no circularity is identified.

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

The paper makes no theoretical derivation; its central claim rests on empirical assumptions about masking schedules, sentinel values, and pretrained VAE suitability. Hyperparameters such as mask probability and schedule endpoints are hand-chosen or tuned, so they are free parameters rather than derived quantities.

free parameters (3)
  • Conditioning mask probability p_t = constant 0.0 to 0.9 in VAE runs; LDM linear 0.1 to 0.25
    The masking probability per condition is the central hyperparameter of the method and is chosen per experiment, not derived.
  • Sparsity schedule endpoints = e.g., GeoBiked 0.5 to 0.6 increasing; Vehicles 0.6 to 0.5 decreasing
    Start and end sparsity values are selected per dataset without a principled rule and affect reported performance.
  • Architecture hyperparameters = embedding dims 11/19; keypoint dims 203/151; U-Net channels 64/128; batch sizes 140/55/32/128; epochs 393/324
    Tuned via Optuna and grid search; no sensitivity ablations are reported, so the method's robustness to these choices is unknown.
assumptions (5)
  • domain assumption Conditions are missing independently and at random; masking each condition with probability p_t during training is a sufficient simulator for all inference-time sparsity patterns.
    Section 3.2 states 'we apply masking to the conditional information to simulate missing values... we mask each condition with probability p_t.' No structured or correlated missingness is tested.
  • domain assumption All numerical conditions are strictly positive, so the sentinel value -1 cannot collide with a real condition.
    Section 3.2: 'For numerical conditions we set the value to -1, since all numerical variables are strictly positive in our dataset.'
  • domain assumption The pretrained Stable Diffusion 2.1 VAE encoder and decoder preserve enough engineering-domain information for latent diffusion and refinement.
    Section 3.4 and the conclusion note that mcLDM performance is constrained by the pretrained VAE, which is used without adaptation.
  • domain assumption Using an unconditioned encoder with conditioning only in the decoder is preferable to conditioning in the encoder.
    Section 3.3 gives a rationale and cites [8, 30], but the paper provides no ablation comparing this design to a standard conditional encoder.
  • standard math Standard VAE ELBO and diffusion MSE objectives are valid training targets for the proposed architectures.
    Sections 2.2 and 3.3/3.4 invoke the classical VAE and denoising diffusion losses without new derivations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Masked Conditioning for Deep Generative Models." pith.science (2026). https://pith.science/paper/6KODO5U2

@misc{pith2026250516725,
  author       = {Pith},
  title        = {Pith review of: Masked Conditioning for Deep Generative Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6KODO5U2}},
  note         = {Machine review of arXiv:2505.16725}
}
read the original abstract

Datasets in engineering domains are often small, sparsely labeled, and contain numerical as well as categorical conditions. Additionally. computational resources are typically limited in practical applications which hinders the adoption of generative models for engineering tasks. We introduce a novel masked-conditioning approach, that enables generative models to work with sparse, mixed-type data. We mask conditions during training to simulate sparse conditions at inference time. For this purpose, we explore the use of various sparsity schedules that show different strengths and weaknesses. In addition, we introduce a flexible embedding that deals with categorical as well as numerical conditions. We integrate our method into an efficient variational autoencoder as well as a latent diffusion model and demonstrate the applicability of our approach on two engineering-related datasets of 2D point clouds and images. Finally, we show that small models trained on limited data can be coupled with large pretrained foundation models to improve generation quality while retaining the controllability induced by our conditioning scheme.

Figures

Figures reproduced from arXiv: 2505.16725 by the authors.

Figure 1
Figure 1. Left: Architecture of our masked conditioning approach, applied to a VAE. Right: Architecture of our masked conditioning approach applied to a diffusion model. In our model, the latent variable 𝑧 is concatenated with the embedded conditioning vector 𝑒𝑦 forming 𝑧𝑐 = [𝑧; 𝑒𝑦 ] where 𝑧𝑐 ∈ ℝ 𝑑𝑧+𝑑𝑦 . The decoder 𝑝𝜃 (𝑥|𝑧𝑐 ) reconstructs the data, yielding ̂𝑥 ∼ 𝑝𝜃 (𝑥|𝑧𝑐 ). The conditional VAE is trained by maximizing the EL… view at source ↗
Figure 2
Figure 2. Left: MSE for the VAEs and LDMs trained on the GeoBIKED, vehicles and quality checked DVM-Car subset datasets for increasing levels of sparsity in the conditions. The sparsity levels are kept constant for each training run. Right: Mean MSE over sparsity levels over the number of samples in the training dataset (BIKED dataset). P. Mueller et al.: Preprint submitted to Elsevier Page 5 of 12 [PITH_FULL_IMAGE:figures/f… view at source ↗
Figure 3
Figure 3. Qualitative results of reconstructing images from the DVM-Car dataset. The mcLDM is conditioned with the same inputs as the ground truth image is labeled. For the refinement, the mcLDM-generated image is passed to the model as input, together with the prompt. Best viewed when zoomed in. P. Mueller et al.: Preprint submitted to Elsevier Page 8 of 12 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative results of reconstructing images from the GeoBiked dataset. The mcLDM is conditioned with the same inputs as the ground truth image is labeled. For the refinement, the mcLDM-generated image is passed to the model as input, together with the prompt. Best vie…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 20 canonical work pages

  1. [1]

    Optuna: A next-generation hyperparameter optimization framework, in: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining

    Akiba, T., Sano, S., Yanase, T., Ohta, T., Koyama, M., 2019. Optuna: A next-generation hyperparameter optimization framework, in: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining

    work page 2019

  2. [2]

    From Automation to Augmentation: Redefining Engineering Design and Manufacturing in the Age of NextGen-AI

    Alam, M.F., Lentsch, A., Yu, N., Barmack, S., Kim, S., Acemoglu, D., Hart, J., Johnson, S., Ahmed, F., 2024. From Automation to Augmentation: Redefining Engineering Design and Manufacturing in the Age of NextGen-AI. An MIT Exploration of Generative AI doi:10.21428/e4baedd9.e39b392d

    work page doi:10.21428/e4baedd9.e39b392d 2024

  3. [3]

    Curriculum learning, in: Proceedings of the 26th Annual International Conference on Machine Learning, ACM, Montreal Quebec Canada

    Bengio, Y., Louradour, J., Collobert, R., Weston, J., 2009. Curriculum learning, in: Proceedings of the 26th Annual International Conference on Machine Learning, ACM, Montreal Quebec Canada. pp. 41–48. URL:https://dl.acm.org/doi/10.1145/1553374.1553380, doi:10.1145/1553374.1553380

    arXiv 2009

  4. [4]

    Understanding and Improving Interpolation in Autoencoders via an Adversarial Regularizer

    Berthelot, D., Raffel, C., Roy, A., Goodfellow, I., 2018. Understanding and Improving Interpolation in Autoencoders via an Adversarial Regularizer. arXiv:1807.07543

    arXiv 2018

  5. [5]

    Burnap,A.,Liu,Y.,Pan,Y.,Lee,H.,Gonzalez,R.,Papalambros,P.Y.,2016. EstimatingandExploringtheProductFormDesignSpaceUsing DeepGenerativeModels,in:Volume2A:42ndDesignAutomationConference,AmericanSocietyofMechanicalEngineers,Charlotte,North Carolina, USA. p. V02AT03A013. doi:10.1115/DETC2016-60091

    work page doi:10.1115/detc2016-60091 2016

  6. [6]

    PaDGAN: Learning to Generate High-Quality Novel Designs

    Chen, W., Ahmed, F., 2021. PaDGAN: Learning to Generate High-Quality Novel Designs. Journal of Mechanical Design 143, 031703. doi:10.1115/1.4048626

    doi:10.1115/1.4048626 2021

  7. [7]

    Mo-padgan: Reparameterizing engineering designs for augmented multi-objective optimization

    Chen, W., Ahmed, F., Nobari, A., 2021. Mo-padgan: Reparameterizing engineering designs for augmented multi-objective optimization. Applied Soft Computing 113, 107909. doi:10.1016/j.asoc.2021.107909, arXiv:2009.07110

    arXiv 2021

  8. [8]

    Image Super-Resolution With Deep Variational Autoencoders

    Chira, D., Haralampiev, I., Winther, O., Dittadi, A., Liévin, V., 2022. Image super-resolution with deep variational autoencoders. URL: https://arxiv.org/abs/2203.09445, arXiv:2203.09445

    work page Pith review arXiv 2022

Show all 49 references
  1. [9]

    Vaes in the presence of missing data

    Collier, M., Nazabal, A., Williams, C.K.I., 2021. Vaes in the presence of missing data. URL:https://arxiv.org/abs/2006.05301, arXiv:2006.05301

    2021 arXiv

  2. [10]

    Diffusionmodelsbeatgansonimagesynthesis,in:Proceedingsofthe35thInternationalConferenceonNeural Information Processing Systems, Curran Associates Inc., Red Hook, NY, USA

    Dhariwal,P.,Nichol,A.,2021. Diffusionmodelsbeatgansonimagesynthesis,in:Proceedingsofthe35thInternationalConferenceonNeural Information Processing Systems, Curran Associates Inc., Red Hook, NY, USA

    2021

  3. [11]

    Scaling rectified flow transformers for high-resolution image synthesis

    Esser, P., Kulal, S., Blattmann, A., Entezari, R., Müller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., Podell, D., Dockhorn, T., English, Z., Lacey, K., Goodwin, A., Marek, Y., Rombach, R., 2024. Scaling rectified flow transformers for high-resolution image syn...

    2024 arXiv

  4. [12]

    Plantldm: A latent diffusion model for visual synthesis of plant images

    Fischer, J., Schaeffler, S., 2022. Plantldm: A latent diffusion model for visual synthesis of plant images. https://github.com/ joh-fischer/PlantLDM

    2022

  5. [13]

    Clipscore: A reference-free evaluation metric for image captioning

    Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y., 2022. Clipscore: A reference-free evaluation metric for image captioning. URL: https://arxiv.org/abs/2104.08718, arXiv:2104.08718

    2022 arXiv

  6. [14]

    Denoising diffusion probabilistic models

    Ho, J., Jain, A., Abbeel, P., 2020. Denoising diffusion probabilistic models. URL: https://arxiv.org/abs/2006.11239, arXiv:2006.11239

    2020 arXiv

  7. [15]

    Dvm-car: A large-scale automotive dataset for visual marketing research and applications, in: Proceedings of IEEE International Conference on Big Data, pp

    Huang, J., Chen, B., Luo, L., Yue, S., , Ounis, I., 2022. Dvm-car: A large-scale automotive dataset for visual marketing research and applications, in: Proceedings of IEEE International Conference on Big Data, pp. 4130–4137

    2022

  8. [16]

    Variational autoencoder with arbitrary conditioning, in: International Conference on Learning Representations

    Ivanov, O., Figurnov, M., Vetrov, D., 2019. Variational autoencoder with arbitrary conditioning, in: International Conference on Learning Representations. URL: https://openreview.net/forum?id=SyxtJh0qYm. P. Mueller et al.:Preprint submitted to Elsevier Page 10 of 12 Masked Con...

    2019

  9. [17]

    Alias-free generative adversarial networks

    Karras, T., Aittala, M., Laine, S., Härkönen, E., Hellsten, J., Lehtinen, J., Aila, T., 2021. Alias-free generative adversarial networks. URL: https://arxiv.org/abs/2106.12423, arXiv:2106.12423

    2021 arXiv

  10. [18]

    Analyzing and improving the image quality of stylegan

    Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T., 2020. Analyzing and improving the image quality of stylegan. URL: https://arxiv.org/abs/1912.04958, arXiv:1912.04958

    2020 arXiv

  11. [19]

    Auto-Encoding Variational Bayes, in: ICLR 2014, arXiv.arXiv:1312.6114

    Kingma, D., Welling, M., 2013. Auto-Encoding Variational Bayes, in: ICLR 2014, arXiv.arXiv:1312.6114

    2013 arXiv

  12. [20]

    An introduction to variational autoencoders

    Kingma, D.P., Welling, M., 2019. An introduction to variational autoencoders. URL: http://dx.doi.org/10.1561/2200000056, doi:10.1561/2200000056

    2019 doi

  13. [21]

    Flux.https://github.com/black-forest-labs/flux

    Labs, B.F., 2023. Flux.https://github.com/black-forest-labs/flux

    2023

  14. [22]

    Vaem: a deep generative model for heterogeneous mixed type data, in: 34th Conference on Neural Information Processing Systems (NeurIPS 2020)

    Ma, C., Tschiatschek, S., Hernández-Lobato, J.M., Turner, R., Zhang, C., 2020. Vaem: a deep generative model for heterogeneous mixed type data, in: 34th Conference on Neural Information Processing Systems (NeurIPS 2020). URL:https://arxiv.org/abs/2006.11941, arXiv:2006.11941

    2020 arXiv

  15. [23]

    Exploring the potentials and challenges of deep generative models in product design conception

    Mueller, P., Mikelsons, L., 2024. Exploring the potentials and challenges of deep generative models in product design conception. arXiv:2407.11104

    2024 arXiv

  16. [24]

    Geobiked:Adatasetwithgeometricfeaturesandautomatedlabelingtechniquestoenabledeep generative models in engineering design

    Mueller,P.,Mueller,S.,Mikelsons,L.,2024a. Geobiked:Adatasetwithgeometricfeaturesandautomatedlabelingtechniquestoenabledeep generative models in engineering design. URL:https://arxiv.org/abs/2409.17045, arXiv:arXiv:2409.17045

    arXiv

  17. [25]

    Insertdiffusion: Identity preserving visualization of objects through a training-free diffusion architecture.arXiv:2407.10592

    Mueller, P., Wiese, J., Craciun, I., Mikelsons, L., 2024b. Insertdiffusion: Identity preserving visualization of objects through a training-free diffusion architecture.arXiv:2407.10592

    arXiv

  18. [26]

    Handling incomplete heterogeneous data using vaes

    Nazábal, A., Olmos, P.M., Ghahramani, Z., Valera, I., 2020. Handling incomplete heterogeneous data using vaes. Pattern Recognition 107, 107501.URL: https://www.sciencedirect.com/science/article/pii/S0031320320303046,doi: https://doi.org/10.1016/ j.patcog.2020.107501

    2020

  19. [27]

    PcDGAN:AContinuousConditionalDiverseGenerativeAdversarialNetworkForInverseDesign, in: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pp

    Nobari,A.H.,Chen,W.,Ahmed,F.,2021a. PcDGAN:AContinuousConditionalDiverseGenerativeAdversarialNetworkForInverseDesign, in: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pp. 606–616. doi:10.1145/3447548. 3467414, arXiv:2106.03620

    arXiv

  20. [28]

    CreativeGAN: Editing Generative Adversarial Networks for Creative Design Synthesis

    Nobari, A.H., Rashad, M.F., Ahmed, F., 2021b. CreativeGAN: Editing Generative Adversarial Networks for Creative Design Synthesis. arXiv:2103.06242

    arXiv

  21. [29]

    Semi-supervised learning with generative adversarial networks

    Odena, A., 2016. Semi-supervised learning with generative adversarial networks. URL: https://arxiv.org/abs/1606.01583, arXiv:1606.01583

    2016 arXiv

  22. [30]

    Conditionalimagegenerationwithpixelcnn decoders

    vandenOord,A.,Kalchbrenner,N.,Vinyals,O.,Espeholt,L.,Graves,A.,Kavukcuoglu,K.,2016. Conditionalimagegenerationwithpixelcnn decoders. URL: https://arxiv.org/abs/1606.05328, arXiv:1606.05328

    2016 arXiv

  23. [31]

    Scalable Diffusion Models with Transformers

    Peebles, W., Xie, S., 2023. Scalable Diffusion Models with Transformers. doi:10.48550/arXiv.2212.09748

    2023 doi arXiv

  24. [32]

    From Concept to Manufacturing: Evaluating Vision-Language Models for Engineering Design.arXiv:2311.12668

    Picard, C., Edwards, K.M., Doris, A.C., Man, B., Giannone, G., Alam, M.F., Ahmed, F., 2023. From Concept to Manufacturing: Evaluating Vision-Language Models for Engineering Design.arXiv:2311.12668

    2023 arXiv

  25. [33]

    State of the Art on Diffusion Models for Visual Computing

    Po,R.,Yifan,W.,Golyanik,V.,Aberman,K.,Barron,J.T.,Bermano,A.H.,Chan,E.R.,Dekel,T.,Holynski,A.,Kanazawa,A.,Liu,C.K.,Liu, L., Mildenhall, B., Nießner, M., Ommer, B., Theobalt, C., Wonka, P., Wetzstein, G., 2023. State of the Art on Diffusion Models for Visual Computing. arXiv:2310.07204

    2023 arXiv

  26. [34]

    SDXL:ImprovingLatentDiffusion Models for High-Resolution Image Synthesis.arXiv:2307.01952

    Podell,D.,English,Z.,Lacey,K.,Blattmann,A.,Dockhorn,T.,Müller,J.,Penna,J.,Rombach,R.,2023. SDXL:ImprovingLatentDiffusion Models for High-Resolution Image Synthesis.arXiv:2307.01952

    2023 arXiv

  27. [35]

    Understanding Deep Learning

    Prince, S., 2023. Understanding Deep Learning. MIT Press

    2023

  28. [36]

    BIKED: A Dataset for Computational Bicycle Design with Machine Learning Benchmarks

    Regenwetter, L., Curry, B., Ahmed, F., 2021. BIKED: A Dataset for Computational Bicycle Design with Machine Learning Benchmarks. arXiv:2103.05844

    2021 arXiv

  29. [37]

    High-Resolution Image Synthesis with Latent Diffusion Models, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, arXiv.arXiv:2112.10752

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B., 2022. High-Resolution Image Synthesis with Latent Diffusion Models, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, arXiv.arXiv:2112.10752

    2022 arXiv

  30. [38]

    U-net: Convolutional networks for biomedical image segmentation

    Ronneberger, O., Fischer, P., Brox, T., 2015. U-net: Convolutional networks for biomedical image segmentation. Medical Image Computing andComputer-AssistedIntervention–MICCAI2015abs/1505.04597. URL: https://api.semanticscholar.org/CorpusID:3719281

    2015 arXiv

  31. [39]

    Palette: Image-to-image diffusion models, in: ACM SIGGRAPH 2022 Conference Proceedings, Association for Computing Machinery, New York, NY, USA

    Saharia, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M., 2022. Palette: Image-to-image diffusion models, in: ACM SIGGRAPH 2022 Conference Proceedings, Association for Computing Machinery, New York, NY, USA. URL:https://doi.org/ 10.1145/3528233.3...

    2022

  32. [40]

    Laion-5b: An open large-scale dataset for training next generation image-text models

    Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., Jitsev, J., 2022. Laion-5b: An open large-scale dataset for training next ...

    2022 arXiv

  33. [41]

    Sohn, K., Yan, X., Lee, H., 2015. Learning structured output representation using deep conditional generative models, in: Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 2, MIT Press, Cambridge, MA, USA. p. 3483–3491

    2015

  34. [42]

    Denoising diffusion implicit models

    Song, J., Meng, C., Ermon, S., 2022. Denoising diffusion implicit models. URL: https://arxiv.org/abs/2010.02502, arXiv:2010.02502

    2022 arXiv

  35. [43]

    Springenberg, J.T., 2016. Unsupervised and semi-supervised learning with categorical generative adversarial networks, in: In pro- ceedings of the Internationcal Conference of Learning Representations (ICLR 2016). URL:https://arxiv.org/abs/1511.06390, arXiv:1511.06390

    2016 arXiv

  36. [44]

    Attention is all you need, in: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.u., Polosukhin, I., 2017. Attention is all you need, in: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (Eds.), Advances in Neural Information P...

    2017

  37. [45]

    Diffusers: State-of-the-art diffusion models

    von Platen, P., Suraj, P., Lozhkov, A., Cuenca, P., Lambert, N., Rasul, K., Davaadorj, M., 2024. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers. P. Mueller et al.:Preprint submitted to Elsevier Page 11 of 12 Masked Conditioning for Deep ...

    2024

  38. [46]

    Image quality assessment: from error visibility to structural similarity

    Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E., 2004. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing 13, 600–612. doi:10.1109/TIP.2003.819861

    2004

  39. [47]

    Generating various airfoil shapes with required lift coefficient using conditional variational autoencoders

    Yonekura, K., Wada, K., Suzuki, K., 2021. Generating various airfoil shapes with required lift coefficient using conditional variational autoencoders. arXiv:2106.09901

    2021 arXiv

  40. [48]

    The unreasonable effectiveness of deep features as a perceptual metric

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O., 2018. The unreasonable effectiveness of deep features as a perceptual metric. URL: https://arxiv.org/abs/1801.03924, arXiv:1801.03924

    2018 arXiv

  41. [49]

    3D Shape Synthesis for Conceptual Design and Optimization Using Variational Autoencoders.arXiv:1904.07964

    Zhang, W., Yang, Z., Jiang, H., Nigam, S., Yamakawa, S., Furuhata, T., Shimada, K., Kara, L.B., 2019. 3D Shape Synthesis for Conceptual Design and Optimization Using Variational Autoencoders.arXiv:1904.07964. P. Mueller et al.:Preprint submitted to Elsevier Page 12 of 12

    2019 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.