Pith. sign in

REVIEW 3 major objections 6 minor 50 references

GUST: Quantifying Free-Form Geometric Uncertainty of Metamaterials Using Small Data

T0 review · 3 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 A pretrained diffusion model learns real 3D-print geometry uncertainty from about a thousand cells.

desk verdict A useful transfer-learning recipe for geometric UQ, with real-world evidence that's promising but not yet quantitatively established. read the letter →

arxiv 2506.12051 v2 pith:A6SG73QJ submitted 2025-05-28 cs.LG cs.CE

Jiahui Zheng , Cole Jahnke , Wei "Wayne" Chen This is my paper
classification cs.LGcs.CE
keywords geometricuncertaintyquantificationmetamaterialsconditionaldiffusionmodelstransferlearningself-supervisedpretrainingfree-formdeformationadditivemanufacturinghomogenizedelasticproperties
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

GUST asks whether a deep generative model can quantify manufacturing-induced geometric uncertainty in metamaterials when real data are scarce. The paper's claim is yes: a conditional diffusion model pretrained on large-scale synthetic pairs of nominal and free-form-deformed geometries, then fine-tuned on about a thousand real 3D-printed unit cells, learns the conditional distribution $p(x_{\mathrm{fab}} \mid x_{\mathrm{nom}})$ of as-fabricated geometry given a nominal design. Across a synthetic experiment and a real stereolithography experiment, GUST's generated geometries and their homogenized elastic-tensor distributions match ground truth better than direct training, dilation-erosion, and Gaussian random field baselines on density, coverage, and Wasserstein distance. If this holds, free-form geometric uncertainty quantification becomes affordable in settings where collecting fabricated samples is the bottleneck, with downstream value for designing aerospace and biomedical components.

What carries the argument

The load-bearing object is a conditional DDPM with a U-Net backbone that takes the noisy as-fabricated geometry $x_{\mathrm{fab},t}$ together with the nominal geometry $x_{\mathrm{nom}}$ as a spatial condition, injected by input concatenation and by SPADE (spatially-adaptive normalization) layers, and is trained to predict the noise added in the forward diffusion process. The two-stage recipe carries the argument: first, self-supervised pretraining on synthetic pairs $(x_{\mathrm{nom}}, x_{\mathrm{fab}})$ generated by stochastic geometric operators, free-form deformation with Gaussian control-point perturbations and a random-hole operator, learns the structure of geometric variability; second, transfer-learning fine-tuning on small real pairs shifts the conditional distribution toward the actual manufacturing process. Monte Carlo sampling from the fine-tuned model then yields as-fabricated geometries whose homogenized elastic properties quantify property uncertainty.

What would settle it

Falsify by applying GUST to a real process whose dominant defects cannot be produced by free-form deformation or random-hole operators, for example anisotropic shrinkage along the build direction, delamination, or support-material remnants, under the same small-data budget; the central claim fails if the fine-tuned model's homogenized property distributions are no closer to ground truth than direct training, or if generated samples retain source-domain hole patterns absent from the real parts.

Watch

Extended reading notes

Core claim

The central claim is that GUST, a conditional denoising diffusion model, can learn the full high-dimensional conditional distribution of as-fabricated unit-cell geometries given a nominal design without the large real datasets normally required. The paper argues this by first pretraining the model on 60,000 synthetic geometries obtained by applying free-form deformation (FFD) with control-point noise $\sigma=6$ to 3,000 nominal designs, then fine-tuning on a small real dataset of stereolithography-printed unit cells (about a thousand retained cells from 17 designs, printed in two passes). Across two experiments, a synthetic target with FFD $\sigma=13$ plus random holes and a real SLA target, GUST's generated geometries and their homogenized elastic-tensor components $C_{11}, C_{12}, C_{22}, C_{33}$ are closer to ground truth than direct training, dilation-erosion with maximum-likelihood scales, and Gaussian random field baselines, as measured by density, coverage, and Wasserstein distance. The paper takes this as evidence that the synthetic-to-real transfer recipe makes free-form geometric UQ feasible with small real data.

Load-bearing premise

The load-bearing premise is that the hand-chosen synthetic pretraining distribution, free-form deformation with control-point noise $\sigma=6$ and no holes, lies close enough to a real manufacturing process that fine-tuning on about a thousand real cells moves the conditional distribution to the true one instead of anchoring it near the source.

Editorial extensions

If this is right

  • With roughly a thousand real unit cells, GUST beats direct training, dilation-erosion, and Gaussian random field baselines on geometry and property metrics, making generative geometric UQ practical where physical fabrication and imaging limit data collection.
  • The learned conditional distribution $p_\theta(x_{\mathrm{fab}} \mid x_{\mathrm{nom}})$ enables Monte Carlo sampling for design optimization under uncertainty, using homogenization to map geometry samples to elastic-property statistics.
  • The near-identical results across transfer-learning freezing strategies suggest the fine-tuning step is insensitive to which U-Net layers are transferred, so the pretraining distribution rather than layer choice is the main determinant of success.
  • Augmenting the small real dataset directly narrows the learned distribution, whereas GUST's pretraining-then-fine-tuning sequence preserves closer alignment with ground truth.
  • Because the recipe is modular, replacing the image-conditioned U-Net with point-cloud, voxel, or mesh diffusion extends the same pretrain-fine-tune-sample pipeline to 3D and vectorized representations.

Reading between the lines

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

  • Editorial inference: if transfer-strategy insensitivity is general, the key design choice in applying GUST to a new process is the pretraining operator family, so the paper's recipe invites a controlled study of how $\sigma$, hole statistics, and operator diversity affect convergence after fine-tuning.
  • Editorial inference: for processes whose dominant defects are outside the FFD-plus-holes family, such as anisotropic shrinkage, delamination, or support-material artifacts, the pretraining prior could mislead; users should tailor the synthetic operator set to known defect physics before trusting the fine-tuned distribution.
  • Editorial inference: the paper's geometric metrics are pixel-distance based and sensitive to misalignment, so a topology- or signed-distance-aware evaluation might make real-world geometry comparisons more decisive than density and coverage in raw pixel space.
  • Editorial inference: the same conditional generative recipe could be tested on wear or erosion of aerodynamic surfaces or microfluidic devices, with the synthetic pretraining operators chosen to mimic the relevant degradation mechanisms and a small field dataset for fine-tuning.
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

3 major / 6 minor

Summary. The paper proposes GUST, a conditional diffusion model that first pretrains on a large synthetic dataset of unit-cell geometries perturbed by free-form deformation (FFD) and random hole nucleation, then fine-tunes on a small set of real stereolithography (SLA) fabricated unit cells. The goal is to quantify the full conditional distribution of as-fabricated geometry given a nominal design, and to propagate that uncertainty to homogenized elastic properties. GUST is compared with three baselines (direct training, dilation-erosion, Gaussian random field) plus an augmented direct training baseline in the appendix. The paper reports quantitative gains for GUST in a synthetic experiment (density, coverage, Wasserstein distances over 30 designs) and qualitative gains in a real-world experiment with two test designs.

Significance. If the central claim is substantiated, GUST is a meaningful contribution: it combines self-supervised synthetic pretraining with transfer learning to reduce real-world data requirements for free-form geometric UQ, and it provides a new real fabrication dataset (SLA-printed unit cells) that is valuable to the community. The synthetic experiment is extensive, uses multiple eval designs and samples, and shows clear quantitative advantages over direct training and GRF baselines. The authors are also transparent about several limitations, including the difficulty of computing density/coverage on real data and the source-target gap. However, the evidence for the real-world claim is currently weaker than the abstract and conclusion suggest, and the geometric evaluation metric has a free parameter that is not fully robust.

major comments (3)
  1. [Sec. 5.2, Figs. 13 and 15] The central real-world claim is not quantitatively supported. The paper states that density and coverage cannot be reliably computed on real samples (Sec. 5.2, paragraph 3), and no alternative geometry distance is reported. The property comparison is limited to two test designs with no Wasserstein distances, confidence intervals, or significance tests, so the statement that "the property distribution produced by GUST most closely approximates the ground truth compared to the three baselines" rests on visual KDE inspection. With only two test designs, this could be chance; the real-world evaluation needs either a quantitative property metric with bootstrap intervals, a robust geometric distance, or at minimum an explicit caveat that the real-world conclusion is qualitative.
  2. [Sec. 4.5, Fig. 10] The density and coverage metrics are computed in a 3D t-SNE embedding whose perplexity was selected post hoc (perplexity 10) because it "led to the most reasonable representation" of the data. This introduces a tunable parameter directly into the evaluation metric. No sensitivity analysis over perplexity is provided, and no independent criterion is given for why this embedding is the right space for computing neighborhood-based fidelity and diversity metrics. The synthetic geometry comparison (Fig. 10) should report metrics across a range of perplexities, or justify the choice with a principled procedure.
  3. [Sec. 5.1 and Sec. 4.1] The synthetic experiment is partly self-referential: pretraining data (FFD with sigma=6), fine-tuning data (FFD with sigma=13 plus random holes), and ground truth are all generated by the same family of geometric operators. This demonstrates that GUST can adapt within a known transformation family, but it does not test adaptation to a qualitatively different manufacturing process. The paper acknowledges in Sec. 5.1 that source-target closeness facilitates transfer, and it correctly uses the real-world experiment as the out-of-family test; however, since that real-world experiment lacks quantitative metrics (see major comment above), the generalizability claim is not yet established. A concrete test would be to apply GUST to a distinct process (or to synthetic data generated by a different perturbation family, such as anisotropic shrinkage) and measure whether the pretraining prior helps or hurts.
minor comments (6)
  1. [Abstract vs. Sec. 4.1] The abstract says 960 unit cells were additively manufactured, while Sec. 4.1 states that 17 designs with 64 unit cells each (1,088 cells) were retained after filtering. Please reconcile these numbers and clarify whether the 960 figure refers to an earlier stage of the pipeline.
  2. [Sec. 4.4, GRF baseline] The GRF baseline hyperparameters (sigma, ell_1, ell_2, and truncation rank M) are not reported or fitted. Since the comparison depends on the strength of the GRF perturbation, please specify the values used and, ideally, a sensitivity study.
  3. [Sec. 5.1, Table 1] The p-values in Table 1 are reported without stating which statistical test was used or whether the comparisons are paired across the 30 nominal designs. Please name the test (e.g., Welch's t-test) and report effect sizes or confidence intervals.
  4. [Figures 8 and 13] The KDE plots are based on a limited number of ground-truth samples (64 per design in the real experiment) and generated samples; the captions do not state sample sizes or kernel bandwidths. Please add these details so the visual comparisons can be interpreted quantitatively.
  5. [Sec. 4.4 and Appendix D] The Augmented DT baseline is introduced only in Appendix D but is a natural alternative to GUST; consider moving it to the main text (or at least referencing it in Sec. 4.4) because it directly addresses the question of whether synthetic generation is better used as pretraining versus as data augmentation.
  6. [Eq. (6)] The notation "FFFD" in Eq. (6) is inconsistent with the description "FFD" elsewhere; please unify the symbol. Also, the left-hand side uses x but the right-hand side only depends on parametric coordinates (u,v) and control points, so the definition of the deformed geometry should be stated more explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: GUST's synthetic pretraining, real-data fine-tuning, and held-out evaluation are distinct, and the weak real-world evidence is a support gap, not a definitional reduction.

full rationale

GUST's derivation chain is not circular. The model is pretrained on synthetic FFD-deformed geometries (Sec 4.1), fine-tuned on either a synthetic or a real as-fabricated dataset, and evaluated against held-out ground-truth distributions using density, coverage, and Wasserstein distance (Sec 4.5, 5.1). No parameter entering the central claim is fitted to the test data and then reported as a prediction: the pretraining noise level (sigma=6) is fixed before real-data fine-tuning, and the MLE-fitted dilation/erosion scales are explicitly presented as point-estimate baselines rather than GUST predictions. The load-bearing real-world experiment uses independently printed unit cells from 17 designs not contained in the pretraining data (Sec 4.1, 5.2). The genuine weakness is evaluation support, not circularity: Sec 5.2 states that density and coverage 'cannot be reliably computed' on real samples due to misalignment sensitivity, and property KDE comparisons are shown for only two test designs, so the real-world claim of closest alignment rests partly on visual inspection. That is an evidence limitation and a correctness risk, not a definitional equivalence. The paper also acknowledges the remaining source-target discrepancy and calls for larger, more diverse pretraining data (Sec 6). The one mildly self-referential aspect is that the synthetic validation target is generated from the same FFD/hole operator family used to create the pretraining data (Sec 4.1), making that part of the demonstration a controlled transfer benchmark within a hand-coded transformation family; however, it does not force the real-world result by construction, and no load-bearing self-citation chain is present.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a small number of hand-chosen data-generation parameters and domain assumptions: the pretraining perturbation scale, the t-SNE evaluation hyperparameter, and the assumption that the synthetic transformation family is a useful prior for real manufacturing variability. No invented physical entities are introduced.

free parameters (7)
  • pretraining_FFD_sigma = 6
    Standard deviation of Gaussian perturbations to FFD control points used to create D_pretrain (Sec 4.1). Hand-chosen; controls the amount of synthetic variability the model sees before fine-tuning.
  • FFD_control_lattice_size_m = Not reported
    The m x m control lattice in Eq. 6 is not specified in the experiments, leaving the pretraining perturbation scale incompletely defined.
  • synthetic_finetune_FFD_sigma = 13
    Standard deviation used for FFD in D_finetune,synth (Sec 4.1); chosen to create a target domain different from pretraining.
  • random_hole_parameters = Not fully specified
    Hole nucleation uses sampled means, covariances, and scales (Eq. 7); the distributions of these quantities are described only generically and the actual values used in the experiments are not given (Sec 3.1, 4.1).
  • tSNE_perplexity = 10
    Perplexity for the t-SNE projection used in density and coverage metrics; chosen post hoc as the 'most reasonable representation' (Sec 4.5), affecting the reported geometry metrics.
  • GRF_baseline_parameters = Not reported
    The GRF baseline requires correlation lengths, variance, and KL truncation order; none are stated (Sec 4.4), making the baseline comparison not fully reproducible.
  • dilation_erosion_scales = alpha=2,beta=1 (synthetic); alpha=5,beta=1 (real)
    Kernel sizes for the dilation-erosion baseline fitted by maximum likelihood estimation (Sec 4.4, 5.1, 5.2). These are fitted to the target data, so the baseline is tuned to the test distribution.
assumptions (5)
  • domain assumption Manufacturing uncertainty in metamaterial unit cells can be represented as a conditional distribution p(x_fab | x_nom) over 64x64 binary pixel images.
    Central modeling choice in Sec 3.2; all geometric variability, including topological changes, must be captured in this fixed-resolution 2D representation.
  • ad hoc to paper The synthetic pretraining distribution generated by FFD and hole nucleation provides a useful prior for real manufacturing uncertainty.
    Assumed in Sec 4.1 and acknowledged in Sec 5.1 where the authors state that performance depends on source-target similarity; if real defects fall outside this transformation family, transfer learning would not help.
  • domain assumption The excluded unit cells (non-penetrating holes) and the retained inner 3x3 grid are not needed to characterize the uncertainty distribution.
    Data filtering in Sec 4.1 removes cells that the imaging pipeline cannot capture; the paper gives no evidence that excluded failure modes are negligible for UQ.
  • standard math The homogenization procedure [38] accurately maps binary geometry to effective elastic tensor components.
    Used in Sec 4.5 to compare property distributions; standard in the metamaterials literature.
  • domain assumption Density and coverage computed in a 3D t-SNE embedding preserve the neighbor structure needed for k-NN based metrics.
    Sec 4.5 uses t-SNE with perplexity 10 to reduce dimensionality; t-SNE is stochastic and distance-distorting, so this assumption is fragile.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GUST: Quantifying Free-Form Geometric Uncertainty of Metamaterials Using Small Data." pith.science (2026). https://pith.science/paper/A6SG73QJ

@misc{pith2026250612051,
  author       = {Pith},
  title        = {Pith review of: GUST: Quantifying Free-Form Geometric Uncertainty of Metamaterials Using Small Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A6SG73QJ}},
  note         = {Machine review of arXiv:2506.12051}
}
read the original abstract

This paper introduces GUST (Generative Uncertainty learning via Self-supervised pretraining and Transfer learning), a framework for quantifying free-form geometric uncertainties inherent in the manufacturing of metamaterials. GUST leverages the representational power of deep generative models to learn a high-dimensional conditional distribution of as-fabricated unit cell geometries given nominal designs, thereby enabling uncertainty quantification. To address the scarcity of real-world manufacturing data, GUST employs a two-stage learning process. First, it leverages self-supervised pretraining on a large-scale synthetic dataset to capture the structure variability inherent in metamaterial geometries and an approximated distribution of as-fabricated geometries given nominal designs. Subsequently, GUST employs transfer learning by fine-tuning the pretrained model on limited real-world manufacturing data, allowing it to adapt to specific manufacturing processes and nominal designs. With only 960 unit cells additively manufactured in only two passes, GUST can capture the variability in geometry and effective material properties. In contrast, directly training a generative model on the same amount of real-world data proves insufficient, as demonstrated through both qualitative and quantitative comparisons. This scalable and cost-effective approach significantly reduces data requirements while maintaining the effectiveness in learning complex, real-world geometric uncertainties, offering an affordable method for free-form geometric uncertainty quantification in the manufacturing of metamaterials. The capabilities of GUST hold significant promise for high-precision industries such as aerospace and biomedical engineering, where understanding and mitigating manufacturing uncertainties are critical.

Figures

Figures reproduced from arXiv: 2506.12051 by the authors.

Figure 1
Figure 1. Overview of GUST. We apply various geometric transformations (e.g., free-form deformation) to add shape [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Assumptions on the forms of geometric uncertainties. Prior approaches typically depend on predefined [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The forward diffusion process adds noise to data and the reverse process recovers the original data through [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (14 more)
Figure 4
Figure 4. Figure 4: Manufacturing data collection process converts a nominal binary array geometry to CAD, extrudes and frames [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Examples of 3D printed metamaterial blocks. A total of 180 blocks were printed in two passes using a [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: The architecture of our conditional diffusion model, consisting of a U-Net structure for conditional generation [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Comparison of as-fabricated geometries generated by GRF, direct training (DT), and GUST in the synthetic [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Kernel density estimates (KDEs) of homogenized elastic tensor components ( [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: KDEs of homogenized elastic properties for the synthetic experiment showing the discrepancy between [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Density and coverage (higher is better) metrics for geometries generated by different methods in the synthetic [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Wasserstein distances (lower is better) between property distributions from different approaches and the [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 12
Figure 12. Figure 12: Comparison of as-fabricated geometries generated by GUST and baselines in the real-world experiment. [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]
Figure 13
Figure 13. Figure 13: KDEs of homogenized elastic tensor components ( [PITH_FULL_IMAGE:figures/full_fig_p016_13.png]
Figure 14
Figure 14. Figure 14: Nominal designs in the test data and their nearest neighbors in the fine-tuning data. [PITH_FULL_IMAGE:figures/full_fig_p017_14.png]
Figure 15
Figure 15. Figure 15: KDEs of homogenized elastic properties for the real-world test cases showing the discrepancy between the [PITH_FULL_IMAGE:figures/full_fig_p017_15.png]
Figure 16
Figure 16. Figure 16: KDEs of homogenized elastic tensor components ( [PITH_FULL_IMAGE:figures/full_fig_p021_16.png]
Figure 17
Figure 17. Figure 17: KDEs of homogenized elastic tensor components ( [PITH_FULL_IMAGE:figures/full_fig_p023_17.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 33 canonical work pages

  1. [1]

    Optimization of fused filament fabrication process parameters under uncertainty to maximize part geometry accuracy

    Paromita Nath, Joseph D Olson, Sankaran Mahadevan, and Yung-Tsun Tina Lee. Optimization of fused filament fabrication process parameters under uncertainty to maximize part geometry accuracy. Additive manufacturing, 35:101331, 2020

    work page 2020

  2. [2]

    Haberman, and Carolyn C

    Clinton Morris, Logan Bekker, Michael R. Haberman, and Carolyn C. Seepersad. Design Exploration of Reliably Manufacturable Materials and Structures With Applications to Negative Stiffness Metamaterials and Microstereolithography. Journal of Mechanical Design, 140(11):111415, 09 2018

    work page 2018

  3. [3]

    Haberman

    Tyler Wiest, Carolyn Conner Seepersad, and Michael R. Haberman. Robust design of an asymmetrically absorbing Willis acoustic metasurface subject to manufacturing-induced dimensional variationsa). The Journal of the Acoustical Society of America, 151(1):216–231, 01 2022

    work page 2022

  4. [4]

    Level set based robust shape and topology optimization under random field uncertainties

    Shikui Chen, Wei Chen, and Sang-Hoon Lee. Level set based robust shape and topology optimization under random field uncertainties. Structural and Multidisciplinary Optimization, 41:507–524, 04 2010

    work page 2010

  5. [5]

    A new level-set based approach to shape and topology optimization under geometric uncertainty

    Shikui Chen and Wei Chen. A new level-set based approach to shape and topology optimization under geometric uncertainty. Structural and Multidisciplinary Optimization, 44:1–18, 07 2011

    work page 2011

  6. [6]

    Wang, David Sell, Thaibao Phan, and Jonathan A

    Evan W. Wang, David Sell, Thaibao Phan, and Jonathan A. Fan. Robust design of topology-optimized metasurfaces. Opt. Mater. Express, 9(2):469–482, Feb 2019

    work page 2019

  7. [7]

    Uncertainty quantification of microstructure variability and mechanical behavior of additively manufactured lattice structures

    Nina Korshunova, Iason Papaioannou, Stefan Kollmannsberger, Daninel Straub, and Ernst Rank. Uncertainty quantification of microstructure variability and mechanical behavior of additively manufactured lattice structures. Computer Methods in Applied Mechanics and Engineering, 385:114049, 2021

    work page 2021

  8. [8]

    GAN-DUF: Hierarchical Deep Generative Models for Design Under Free-Form Geometric Uncertainty

    Wei (Wayne) Chen, Doksoo Lee, Oluwaseyi Balogun, and Wei Chen. GAN-DUF: Hierarchical Deep Generative Models for Design Under Free-Form Geometric Uncertainty. Journal of Mechanical Design, 145(1):011703, 10 2022

    work page 2022

Show all 50 references
  1. [9]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and 18 GUST: Quantifying Free-Form Geometric Uncertainty of Metam...

    2014

  2. [10]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6840–6851. Curran Associates, Inc., 2020

    2020

  3. [11]

    Maximum likelihood training of score-based diffusion models

    Yang Song, Conor Durkan, Iain Murray, and Stefano Ermon. Maximum likelihood training of score-based diffusion models. In Neural Information Processing Systems, 2021

    2021

  4. [12]

    Elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Systems, volume 35, pages 26565–...

    2022

  5. [13]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In M. Ranzato, A. Beygelzimer, Y . Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 8780–8794. Curran Associates, Inc., 2021

    2021

  6. [14]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Francis Bach and David Blei, editors, Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings o...

    2015

  7. [15]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674–10685, 2022

    2022

  8. [16]

    Palette: Image-to-image diffusion models

    Chitwan Saharia, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mo- hammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 conference proceedings, pages 1–10, 2022

    2022

  9. [17]

    Cnn features off-the-shelf: an astounding baseline for recognition

    Ali Sharif Razavian, Hossein Azizpour, Josephine Sullivan, and Stefan Carlsson. Cnn features off-the-shelf: an astounding baseline for recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 806–813, 2014

    2014

  10. [18]

    How transferable are features in deep neural networks? Advances in neural information processing systems, 27, 2014

    Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? Advances in neural information processing systems, 27, 2014

    2014

  11. [19]

    Transfer learning for diffusion models

    Yidong Ouyang, Liyan Xie, Hongyuan Zha, and Guang Cheng. Transfer learning for diffusion models. arXiv preprint arXiv:2405.16876, 2024

    2024 arXiv

  12. [20]

    Fine-tuning diffusion models with limited data

    Taehong Moon, Moonseok Choi, Gayoung Lee, Jung-Woo Ha, and Juho Lee. Fine-tuning diffusion models with limited data. In NeurIPS 2022 Workshop on Score-Based Methods, 2022

    2022

  13. [21]

    Difffit: Unlocking transferability of large diffusion models via simple parameter-efficient fine-tuning

    Enze Xie, Lewei Yao, Han Shi, Zhili Liu, Daquan Zhou, Zhaoqiang Liu, Jiawei Li, and Zhenguo Li. Difffit: Unlocking transferability of large diffusion models via simple parameter-efficient fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, ...

    2023

  14. [22]

    One-shot generative domain adaptation

    Ceyuan Yang, Yujun Shen, Zhiyi Zhang, Yinghao Xu, Jiapeng Zhu, Zhirong Wu, and Bolei Zhou. One-shot generative domain adaptation. In Proceedings of the ieee/cvf international conference on computer vision, pages 7733–7742, 2023

    2023

  15. [23]

    Few-shot image generation with diffusion models

    Jingyuan Zhu, Huimin Ma, Jiansheng Chen, and Jian Yuan. Few-shot image generation with diffusion models. arXiv preprint arXiv:2211.03264, 2022

    2022 arXiv

  16. [24]

    Generalized one-shot domain adaptation of generative adversarial networks

    Zicheng Zhang, Yinglu Liu, Congying Han, Tiande Guo, Ting Yao, and Tao Mei. Generalized one-shot domain adaptation of generative adversarial networks. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Syste...

    2022

  17. [25]

    A closer look at few-shot image generation

    Yunqing Zhao, Henghui Ding, Houjing Huang, and Ngai-Man Cheung. A closer look at few-shot image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9140–9150, 2022

    2022

  18. [26]

    Explicit inductive bias for transfer learning with convolutional networks

    LI Xuhong, Yves Grandvalet, and Franck Davoine. Explicit inductive bias for transfer learning with convolutional networks. In International conference on machine learning, pages 2825–2834. PMLR, 2018

    2018

  19. [27]

    Provable sample-efficient transfer learning conditional diffusion models via representation learning

    Ziheng Cheng, Tianyu Xie, Shiyue Zhang, and Cheng Zhang. Provable sample-efficient transfer learning conditional diffusion models via representation learning. arXiv preprint arXiv:2502.04491, 2025

    2025

  20. [28]

    Transfer learning for diffusion models

    Yidong Ouyang, Liyan Xie, Hongyuan Zha, and Guang Cheng. Transfer learning for diffusion models. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors,Advances in Neural Information Processing Systems, volume 37, pages 136962–136989. Cur...

    2024

  21. [29]

    Denoising diffusion probabilistic models for addressing data limitations in chest x-ray classification

    Evi MC Huijben, Josien PW Pluim, and Maureen AJM van Eijnatten. Denoising diffusion probabilistic models for addressing data limitations in chest x-ray classification. Informatics in Medicine Unlocked, 50:101575, 2024

    2024

  22. [30]

    Semantic image synthesis with spatially-adaptive normalization

    Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337–2346, 2019

    2019

  23. [31]

    Mechanical cloak via data- driven aperiodic metamaterial design

    Liwei Wang, Jagannadh Boddapati, Ke Liu, Ping Zhu, Chiara Daraio, and Wei Chen. Mechanical cloak via data- driven aperiodic metamaterial design. Proceedings of the National Academy of Sciences, 119(13):e2122185119, 2022

    2022

  24. [32]

    Freeu: Free lunch in diffusion u-net

    Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4733–4743, June 2024

    2024

  25. [33]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization, 2017

    2017

  26. [34]

    Differential evolution–a simple and efficient heuristic for global optimization over continuous spaces

    Rainer Storn and Kenneth Price. Differential evolution–a simple and efficient heuristic for global optimization over continuous spaces. Journal of global optimization, 11:341–359, 1997

    1997

  27. [35]

    Statistical modeling of images with fields of gaussian scale mixtures

    Siwei Lyu and Eero Simoncelli. Statistical modeling of images with fields of gaussian scale mixtures. Advances in Neural Information Processing Systems, 19, 2006

    2006

  28. [36]

    Improved precision and recall metric for assessing generative models

    Tuomas Kynkäänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information P...

    2019

  29. [37]

    Reliable fidelity and diversity metrics for generative models

    Muhammad Ferjad Naeem, Seong Joon Oh, Youngjung Uh, Yunjey Choi, and Jaejun Yoo. Reliable fidelity and diversity metrics for generative models. In International conference on machine learning , pages 7176–7185. PMLR, 2020

    2020

  30. [38]

    Design of materials using topology optimization and energy-based homogenization approach in matlab

    Liang Xia and Piotr Breitkopf. Design of materials using topology optimization and energy-based homogenization approach in matlab. Structural and multidisciplinary optimization, 52(6):1229–1241, 2015

    2015

  31. [39]

    Optimal transport: old and new, volume 338

    Cédric Villani et al. Optimal transport: old and new, volume 338. Springer, 2009

    2009

  32. [40]

    A scaling law for synthetic-to-real transfer: How much is your pre-training effective? arXiv preprint arXiv:2108.11018, 2021

    Hiroaki Mikami, Kenji Fukumizu, Shogo Murai, Shuji Suzuki, Yuta Kikuchi, Taiji Suzuki, Shin-ichi Maeda, and Kohei Hayashi. A scaling law for synthetic-to-real transfer: How much is your pre-training effective? arXiv preprint arXiv:2108.11018, 2021

    2021 arXiv

  33. [41]

    An empirical study of scaling laws for transfer

    Matthew Barnett. An empirical study of scaling laws for transfer. arXiv preprint arXiv:2408.16947, 2024

    2024 arXiv

  34. [42]

    Scaling laws for transfer

    Danny Hernandez, Jared Kaplan, Tom Henighan, and Sam McCandlish. Scaling laws for transfer. arXiv preprint arXiv:2102.01293, 2021

    2021 arXiv

  35. [43]

    Big transfer (bit): General visual representation learning

    Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, and Neil Houlsby. Big transfer (bit): General visual representation learning. In European conference on computer vision, pages 491–507. Springer, 2020

    2020

  36. [44]

    Revisiting unreasonable effectiveness of data in deep learning era

    Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843–852, 2017

    2017

  37. [45]

    What makes imagenet good for transfer learning? arXiv preprint arXiv:1608.08614, 2016

    Minyoung Huh, Pulkit Agrawal, and Alexei A Efros. What makes imagenet good for transfer learning? arXiv preprint arXiv:1608.08614, 2016

    2016 arXiv

  38. [46]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022

    2022

  39. [47]

    Parameter-efficient fine-tuning of 3d ddpm for mri image generation using tensor networks

    Binghua Li, Ziqing Chang, Tong Liang, Chao Li, Toshihisa Tanaka, Shigeki Aoki, Qibin Zhao, and Zhe Sun. Parameter-efficient fine-tuning of 3d ddpm for mri image generation using tensor networks. arXiv preprint arXiv:2507.18112, 2025

    2025 arXiv

  40. [49]

    One-shot generative domain adaptation

    Ceyuan Yang, Yujun Shen, Zhiyi Zhang, Yinghao Xu, Jiapeng Zhu, Zhirong Wu, and Bolei Zhou. One-shot generative domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7733–7742, October 2023

    2023

  41. [50]

    Image synthesis under limited data: A survey and taxonomy

    Mengping Yang and Zhe Wang. Image synthesis under limited data: A survey and taxonomy. International Journal of Computer Vision, pages 1–38, 2025

    2025

  42. [51]

    as-fabricated

    Karen K Yuen. The two-sample trimmed t for unequal population variances. Biometrika, 61(1):165–170, 1974. 20 GUST: Quantifying Free-Form Geometric Uncertainty of Metamaterials Using Small Data A Additional Results on Property Distributions In this section, we provide more resu...

    1974

Pith tools

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