Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

LAMP: Data-Efficient Linear Affine Weight-Space Models for Parameter-Controlled 3D Shape Generation and Extrapolation

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read LAMP generates parameter-controlled 3D shapes from a few dozen examples by linearly mixing neural shape decoders in weight space.

desk verdict Genuinely useful method paper — LAMP's affine mixing of aligned SDF decoders works empirically on two design datasets, but the main theorem assumes the key geometric step it claims to prove, and the abstract overstates the 50-sample result. read the letter →

arxiv 2510.22491 v3 pith:QGPZOBYP submitted 2025-10-26 cs.LG cs.CEcs.CV

Ghadi Nehme , Yanxia Zhang , Dule Shu , Matt Klenk , Faez Ahmed This is my paper
classification cs.LGcs.CEcs.CV
keywords 3Dshapegenerationsigneddistancefunctionsweight-spaceinterpolationdata-efficientlearningparametricdesignextrapolationneuralimplicitrepresentationssafetymetrics
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

LAMP claims that controllable 3D shape generation can be solved as a linear problem: no large dataset, no learned latent space, no disentangling. The method overfits a signed-distance-function decoder to each of a handful of exemplar meshes, always starting from the same initial weights, so that the weight vectors line up in a common aligned basis. It then picks the linear combination of weights whose implied parameters match the user's target, and decodes the result; the same linear solve works for extrapolation outside the training range and for constrained optimization. The paper adds a lightweight safety metric that measures how far a mix is from the model's local linear regime and shows it flags invalid shapes with high accuracy. If this works, dozens of annotated meshes—not thousands—can become a practical engine for exploring new car or aircraft designs.

What carries the argument

Aligned SDF weight space: each exemplar's shape is encoded in the weights of an SDF MLP overfit from a shared initialization, so all weight vectors live in the same space and are indexed by their parameter vectors. The identity doing the work is the affine mixing rule w_d = Σ α_i w_i with Σ α_i = 1; solving for α under the parameter constraint gives w_d, which decodes to the desired mesh. The local-linearity assumption f(z; Σ α_i w_i) ≈ Σ α_i f(z; w_i) is checked by the linearity-mismatch safety metric, which samples points and compares those two quantities, accepting generation only below a threshold of 0.01.

What would settle it

Take two cars with the same planform but different topology—say a sedan with a roof pillar and a convertible whose roof region is empty—and overfit an SDF decoder to each from the same initialization. Mix them with equal weights. If the linearity-mismatch score at the 0.01 threshold stays low while the surrogate-predicted parameters of the decoded mesh are far from the target or the mesh is collapsed, the claim that low mismatch certifies parameter fidelity is false.

Watch

Extended reading notes

Core claim

The central claim is that a bank of SDF decoders, each overfit to one exemplar from a shared initialization, forms an aligned weight space in which parameter-controlled generation is just an affine mixing problem. Given a target parameter vector and a set of constrained indices, the method solves for coefficients α by minimizing the mismatches on those constraints with a sum-to-one constraint, allowing negative α for extrapolation; the synthesized weights are the weighted average of exemplar weights, and the decoded mesh is the output. Under two assumptions—parameters map linearly to control points, and decoders are locally linear in their weights—the mixed field approximates the signed dist

Load-bearing premise

The proof assumes that a weighted sum of signed distance functions equals the signed distance function of a shape at the weighted control points—an additivity property that is precisely the conclusion the method needs; the paper does not derive it, and if it fails, mixing can pass the safety check yet produce geometry that does not match the target parameters.

Editorial extensions

If this is right

  • With just 50–100 annotated meshes, LAMP can synthesize novel geometries that respect specified design parameters, replacing the need for large training sets in parametric shape generation.
  • Extrapolation up to 100% beyond the dataset range yields plausible meshes, enabling dataset augmentation and exploration of designs outside the training hull.
  • The linearity-mismatch score acts as a data-independent safety certificate, flagging unreliable mixes even when no labeled data are available.
  • Constrained affine mixing can jointly optimize a performance target—like 10% drag reduction—while holding selected physical parameters fixed.

Reading between the lines

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

  • If linear mixing in weight space transfers beyond the two design families tested, the same 'basis of fine-tuned decoders' recipe could apply to other implicit neural fields (density, color, flow) and other parameter-controlled design tasks; this is an extension, not a claim of the paper.
  • Because negative coefficients are permitted, LAMP is doing linear extrapolation in function space; that is likely why it reaches outside the convex hull, but it also means the safety metric is the only guardrail, and the paper tunes its threshold on automotive shapes only—transferring it to new domains needs a check.
  • The surrogate validation with randomly initialized PointNet features achieving high R² hints at a cheap, general-purpose way to audit any generative 3D method's parameter compliance, potentially useful outside LAMP itself.
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. The paper proposes LAMP, a framework for parameter-controlled 3D shape generation that overfits per-exemplar SDF decoders from a shared initialization, solves a constrained affine-mixing problem in weight space, and decodes new meshes. A linearity-mismatch safety metric is introduced to flag invalid mixes. Experiments on DrivAerNet++ and BlendedNet compare LAMP against DNI and AE-LPA for interpolation, large-range extrapolation, and performance-driven optimization, reporting improved Chamfer distance, IoU, parameter error, and R^2 in most settings.

Significance. If the central claims hold, LAMP would be a valuable data-efficient tool for engineering design: it needs only tens to hundreds of exemplars, supports extrapolation beyond the training hull, and provides a human-validated safety check. The paper is also commendable for including two datasets, a surrogate-based validator with R^2>0.9, direct geometric measurement for car length, ablations over sample size, and a human-annotated validation of the safety metric. However, the theoretical justification in Appendix B contains an unproven geometric-additivity step that is central to the claim that affine mixing in weight space produces shapes with the target parameters. The empirical extrapolation results also rely heavily on a surrogate that is trained in-distribution and used beyond its training range. These issues require careful revision before the claims can be accepted.

major comments (4)
  1. [Appendix B, Theorem proof] The proof asserts, without proof or explicit assumption, that 'If the SDFs d_i(z) correspond to shapes with shared topology and smooth variation in geometry, then Σ α_i d_i(z) ≈ SDF(z; C_α).' This is essentially the theorem's conclusion: it is the step that connects linearly mixed SDF values to the SDF of the linearly interpolated control points. It does not follow from assumptions (A1) or (A2), and it is false in general (e.g., non-uniform scaling). The theorem should either be restated with this additivity as an explicit assumption, or proved under concrete conditions on the SDF family. As written, the claim that mixing is 'provably safe' (Section 1) and the central theoretical justification are not supported.
  2. [Section 3, Safety metric] The linearity-mismatch metric measures |f(z; Σ α_i w_i) − Σ α_i f(z; w_i)|, i.e., how close the mixed decoder is to the linear combination of the individual decoders. It does not measure |Σ α_i f(z; w_i) − SDF(z; C_α)|, which is the quantity needed to certify parameter fidelity. A low mismatch score can therefore coexist with geometry that does not correspond to the target parameters. The abstract and introduction claim the metric 'certifies' geometric validity; this is an overstatement. Moreover, the threshold ε=0.01 is selected from the same human-annotated dataset used to validate the metric (Appendix H), and it is unclear whether LAMP's reported results are filtered by this threshold while baseline results are not. Please clarify whether safety-based filtering was applied, and if so, apply the same protocol to all methods.
  3. [Section 4, Large-range extrapolation (Table 3)] For extrapolation beyond the dataset bounds there is no ground-truth mesh. The reported MAE and R^2 are computed using a mesh-based surrogate (Appendix G) that is trained on in-distribution samples and evaluated on held-out in-distribution test sets. Using this surrogate to score targets that are up to 100% beyond the training range assumes the surrogate itself extrapolates reliably, which is not established. Direct geometric measurement is only implemented for car length (Appendix G). The claim that LAMP 'sustains high fidelity (R^2 > 0.86)' in Table 3 is therefore only as strong as the surrogate's out-of-distribution validity. Please provide out-of-distribution validation of the surrogate, or restrict the fidelity claims to settings with direct measurement or ground-truth geometry.
  4. [Abstract vs. Section 4] The abstract states 'as few as 50 samples,' while Section 4 and the conclusion state 'as few as 100 samples.' Table 7 shows that LAMP with 50 samples has R^2 = −0.214 and MAE = 1.083 in the ablation, which does not support the abstract's 50-sample claim. Please align the abstract with the actual reported results and specify the experimental conditions (e.g., number of samples, parameters constrained) for the 50-sample claim.
minor comments (5)
  1. [Tables 1-6] The text refers to 'Table 5' and 'Table 6' for interpolation and extrapolation results, but the corresponding tables are labeled Table 1 and Table 2. Appendix D also contains a Table 6, causing duplicate numbering. Please renumber consistently.
  2. [Appendix C] The proof of local linearity uses a Taylor expansion with a second-order remainder for a ReLU MLP, but ReLU is not twice differentiable. The argument can be repaired by assuming smooth activations (e.g., Softplus) or by using a different local-linearity bound; please revise.
  3. [Figure 3] The caption lists multiple R^2 values without clearly indicating which method and parameter each belongs to. Please add explicit labels or a legend.
  4. [Equation (1)] The notation P^T_{:,C} α is correct only if P is N×d and α is N×1; please define all dimensions explicitly to avoid confusion.
  5. [Tables 1-4] No error bars, confidence intervals, or seeds are reported. Given the small-sample regime and the stochasticity of overfitting and mixing, please provide variance estimates across multiple runs.

Circularity Check

1 steps flagged · score 5.0 of 10

Appendix B's proof assumes the key SDF additivity it must prove; parameter control therefore rests on an unproven geometric postulate, though the experiments are externally evaluated.

  1. self definitional [Appendix B, proof of 'Theoretical Justification of SDF Weight Interpolation' (after defining C_α)]
    "If the SDFs d_i(z) correspond to shapes with shared topology and smooth variation in geometry, then the signed distances satisfy: Σ_i α_i d_i(z) ≈ SDF(z; C_α). Therefore, f(z; ŵα) ≈ SDF(z; C(p̂α))."

    The theorem's conclusion is exactly f(z; ŵα) ≈ SDF(z; C(p̂α)). After assumptions (2) and (3), the proof has reduced the left side to Σ_i α_i d_i(z), and the 'If' sentence asserts the remaining equality Σ_i α_i d_i(z) ≈ SDF(z; C_α). That equality is not derived from (A1) or (A2); it is the geometric additivity of SDFs under control-point blending, which is the substantive fact the theorem claims to establish. The safety metric only checks |f(z; Σα_i w_i) − Σα_i f(z; w_i)|, i.e., weight-space linearity, not |Σα_i d_i(z) − SDF(z; C_α)|, so it cannot certify the missing step.

full rationale

The empirical core of the paper is not fitted circularly: interpolation and extrapolation are tested on held-out meshes with Chamfer/IoU, and parameter compliance is checked by a separately trained surrogate plus direct geometric measurements; baselines are external. The extraction of a safety threshold ε=0.01 from a human-annotated validity set is a calibration, not a circular prediction. However, the theoretical guarantee that weight-space mixing yields a mesh with the target parameters contains a question-begging step: Appendix B assumes, rather than proves, that a weighted sum of SDFs equals the SDF of the linearly blended control-point shape. This is the central link from the mixing solve to geometric parameter control, so the derivation is partially circular even though the empirical results stand independently. Score 5 reflects a central theoretical step that reduces to its own conclusion, with the external evaluation preventing a higher score.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central result rests on domain linearity assumptions (A1, A2), an unproven SDF-additivity step, and a perceptually validated safety threshold. No novel physical entities are introduced.

free parameters (2)
  • Mixing coefficients alpha = per-target least-squares solution
    Solved via Eq. 1 for every target; regularization and underdetermined-case handling not specified.
  • Safety threshold epsilon = 0.01
    Selected from human-annotated ROC/PR analysis in Appendix H; used to accept/reject outputs in all experiments.
assumptions (5)
  • domain assumption Control-point map C(p) is linear in parameters: C(sum alpha_i p_i) = sum alpha_i C(p_i).
    Sec. 3 (A1) and Appendix A; needed to equate mixed parameters with mixed geometry.
  • domain assumption SDF decoder f(z;w) is locally linear in weights: f(z; sum alpha_i w_i) approx sum alpha_i f(z; w_i).
    Sec. 3 (A2) and Appendix C; Taylor bound valid only if trained weights stay near shared init, which extrapolation may violate.
  • domain assumption Each overfit decoder satisfies f(z; w_i) approx SDF(z; C(p_i)).
    Appendix B assumption (2); not quantitatively verified per exemplar.
  • ad hoc to paper Affine combinations of SDFs approximate the SDF of affine-combined control points: sum alpha_i d_i(z) approx SDF(z; C_alpha).
    Appendix B proof inserts this without derivation; it is essentially the theorem's conclusion.
  • domain assumption Linearity mismatch below epsilon implies geometrically valid and parameter-faithful mesh.
    Validated only against human labels for one-parameter sweeps on DrivAerNet++ (Appendix H), not multi-parameter extrapolation or other shape families.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LAMP: Data-Efficient Linear Affine Weight-Space Models for Parameter-Controlled 3D Shape Generation and Extrapolation." pith.science (2026). https://pith.science/paper/QGPZOBYP

@misc{pith2026251022491,
  author       = {Pith},
  title        = {Pith review of: LAMP: Data-Efficient Linear Affine Weight-Space Models for Parameter-Controlled 3D Shape Generation and Extrapolation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QGPZOBYP}},
  note         = {Machine review of arXiv:2510.22491}
}
read the original abstract

Generating high-fidelity 3D geometries under explicit parameter constraints is central to engineering design, yet current methods often require large datasets and fail to provide reliable control beyond the training distribution. We introduce LAMP, a data-efficient framework for controllable and interpretable 3D generation that aligns signed distance function (SDF) decoders by overfitting each exemplar from a shared initialization, then generates new designs by solving a parameter-constrained affine mixing problem in the aligned weight space. To improve reliability, we propose a linearity-mismatch safety metric that detects when mixed decoders leave the valid local regime. We evaluate LAMP on DrivAerNet++, BlendedNet, and additional industry-level vehicle families, including sports cars, SUVs, and convertibles. LAMP enables controlled interpolation with as few as 50 samples, safe extrapolation up to 100% beyond training ranges, and performance-guided optimization under fixed parameters, outperforming conditional autoencoder and Deep Network Interpolation (DNI) baselines in extrapolation, data efficiency, and parameter fidelity. Our results demonstrate that LAMP advances controllable, data-efficient, and safe 3D generation for design exploration, dataset generation, and performance-driven optimization.

Figures

Figures reproduced from arXiv: 2510.22491 by the authors.

Figure 1
Figure 1. Overview of LAMP: (I) aligned SDF weight space construction, (II) parameter-constrained [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Single-parameter extrapolation showing LAMP’s smooth, plausible geometries. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Single-parameter extrapolation beyond the dataset range, with all other parameters allowed [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (17 more)
Figure 4
Figure 4. Figure 4: Four-parameter extrapolation. Left: distribution of generated meshes in a 2D point cloud [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Linearity-mismatch safety metric for diffuser angle extrapolation. Failures (e.g., sample [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Performance-driven drag optimization on DrivAerNet++. Left: target vs. predicted drag [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: LAMP’s R2 scores for single-parameter sweeps on the DrivAerNet++ dataset, extrapo￾lated ±100% beyond the dataset range [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: Single-parameter extrapolation within dataset range on DrivAerNet++. Training samples [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: Single-parameter sweep on BlendedNet. LAMP sustains smooth, plausible geometry [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]
Figure 10
Figure 10. Figure 10: Single-parameter extrapolation beyond the dataset range. All other parameters are al [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]
Figure 11
Figure 11. Figure 11: Safety metric values as a function of target parameter sweeps across six design parame [PITH_FULL_IMAGE:figures/full_fig_p021_11.png]
Figure 12
Figure 12. Figure 12: Diagram of the mesh-based surrogate pipeline. A decoded mesh is first converted into [PITH_FULL_IMAGE:figures/full_fig_p022_12.png]
Figure 13
Figure 13. Figure 13: Predicted vs. ground truth parameters on the DrivAerNet++ test set, evaluating the mesh [PITH_FULL_IMAGE:figures/full_fig_p023_13.png]
Figure 14
Figure 14. Figure 14: Predicted vs. ground truth parameters on the BlendedNet test set, evaluating the mesh [PITH_FULL_IMAGE:figures/full_fig_p024_14.png]
Figure 15
Figure 15. Figure 15: Direct geometric measurement of car length in DrivAerNet++. We slice the decoded [PITH_FULL_IMAGE:figures/full_fig_p025_15.png]
Figure 16
Figure 16. Figure 16: Examples of meshes labeled during human annotation. Left: [PITH_FULL_IMAGE:figures/full_fig_p025_16.png]
Figure 17
Figure 17. Figure 17: Box plot of mean linearity error (log scale) across meshes labeled as valid vs. invalid. [PITH_FULL_IMAGE:figures/full_fig_p026_17.png]
Figure 18
Figure 18. Figure 18: Validation of the linearity-mismatch safety metric against human-annotated mesh va [PITH_FULL_IMAGE:figures/full_fig_p027_18.png]
Figure 19
Figure 19. Figure 19: BlendedNet Parametrization Sung et al. (2025) [PITH_FULL_IMAGE:figures/full_fig_p027_19.png]
Figure 20
Figure 20. Figure 20: DrivAerNet++ Parametrization Elrefaie et al. (2024) [PITH_FULL_IMAGE:figures/full_fig_p028_20.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. CertMix: Certified, Data-Efficient Metamaterial Design by Affine Mixing of Aligned Neural-Implicit Weight Spaces

    cs.LG 2026-07 conditional novelty 6.0 of 10

    Aligned SIREN weight-space affine mixing with a differentiable homogenizer yields certified, data-efficient metamaterial inverse design that extrapolates and beats generative baselines by orders of magnitude.

Reference graph

Works this paper leans on

42 extracted references · 1 canonical work pages · cited by 1 Pith paper

  1. [1]

    Learning representations and generative models for 3d point clouds

    Panos Achlioptas and et al. Learning representations and generative models for 3d point clouds. In ICML, 2018

    2018

  2. [2]

    Learning from randomly initialized neural network features

    Ehsan Amid, Rohan Anil, Wojciech Kot owski, and Manfred K Warmuth. Learning from randomly initialized neural network features. arXiv preprint arXiv:2202.06438, 2022

    arXiv 2022

  3. [3]

    Sal: Sign agnostic learning of shapes from raw data

    Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. CVPR, 2020

    2020

  4. [4]

    Gan cocktail: Mixing GAN s without dataset access

    Omri Avrahami, Dani Lischinski, and Ohad Fried. Gan cocktail: Mixing GAN s without dataset access. In Computer Vision -- ECCV 2022, volume 13683 of Lecture Notes in Computer Science, pp.\ 205--221. Springer, 2022

    2022

  5. [5]

    Diffusion soup: Model merging for text-to-image diffusion models

    Benjamin Biggs, Arjun Seshadri, Yang Zou, Achin Jain, Aditya Golatkar, Yusheng Xie, Alessandro Achille, Ashwin Swaminathan, and Stefano Soatto. Diffusion soup: Model merging for text-to-image diffusion models. In European Conference on Computer Vision (ECCV), 2024

    2024

  6. [6]

    Sdf-diff: Differentiable rendering of signed distance fields for 3d shape editing

    Hang Chen and et al. Sdf-diff: Differentiable rendering of signed distance fields for 3d shape editing. In ECCV, 2022

    2022

  7. [7]

    Maskedgen: Conditional 3d shape generation with masked diffusion

    Yilun Chen and et al. Maskedgen: Conditional 3d shape generation with masked diffusion. 2023

    2023

  8. [8]

    Schwing, and Liang-Yan Gui

    Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G. Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 4456--4465, June 2023

    2023

Show all 42 references
  1. [9]

    Implicit Functions in Feature Space for 3D Shape Reconstruction and Completion

    Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit Functions in Feature Space for 3D Shape Reconstruction and Completion . arXiv.2003.01456, 2020. doi:10.48550/arXiv.2003.01456

    2003 doi arXiv

  2. [10]

    Diffusion-sdf: Conditional generative modeling of signed distance functions

    Gene Chou, Yuval Bahat, and Felix Heide. Diffusion-sdf: Conditional generative modeling of signed distance functions. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp.\ 2262--2272, October 2023

    2023

  3. [11]

    Cad2shape: Learning geometry representations from cad

    Ruida Du and et al. Cad2shape: Learning geometry representations from cad. In ICCV, 2023

    2023

  4. [12]

    Drivaernet++: A large-scale multimodal car dataset with computational fluid dynamics simulations and deep learning benchmarks

    Mohamed Elrefaie, Florin Morar, Angela Dai, and Faez Ahmed. Drivaernet++: A large-scale multimodal car dataset with computational fluid dynamics simulations and deep learning benchmarks. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang (eds....

    2024

  5. [13]

    Hyperdiffusion: Generating implicit neural fields with weight-space diffusion

    Ziya Erko c , Fangchang Ma, Qi Shan, Matthias Nie ner, and Angela Dai. Hyperdiffusion: Generating implicit neural fields with weight-space diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp.\ 14300--14310, October 2023

    2023

  6. [14]

    GET3D : A generative model of high quality 3d textured shapes learned from images

    Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. GET3D : A generative model of high quality 3d textured shapes learned from images. In Advances in Neural Information Processing Systems, 2022

    2022

  7. [15]

    Sdm-net: Deep generative network for structured deformable mesh

    Lin Gao, Jie Yang, Tong Wu, Yu-Jie Yuan, Hongbo Fu, Yu-Kun Lai, and Hao Zhang. Sdm-net: Deep generative network for structured deformable mesh. ACM Transactions on Graphics (TOG), 38 0 (6): 0 1--15, 2019

    2019

  8. [16]

    Atlasnet: A papier-mâché approach to learning 3d surface generation

    Thibault Groueix and et al. Atlasnet: A papier-mâché approach to learning 3d surface generation. In CVPR, 2018

    2018

  9. [17]

    A mechanism for producing aligned latent spaces with autoencoders

    Saachi Jain, Adityanarayanan Radhakrishnan, and Caroline Uhler. A mechanism for producing aligned latent spaces with autoencoders. arXiv preprint arXiv:2106.15456, 2021

    2021 arXiv

  10. [18]

    Salad: Part-level latent diffusion for 3d shape generation and manipulation

    Juil Koo, Seungwoo Yoo, Minh Hieu Nguyen, and Minhyuk Sung. Salad: Part-level latent diffusion for 3d shape generation and manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp.\ 14441--14451, October 2023

    2023

  11. [19]

    Gen-ldm: A generative latent diffusion model for 3d shape generation

    Hao Liu and et al. Gen-ldm: A generative latent diffusion model for 3d shape generation. arXiv preprint arXiv:2301.XXXX, 2023

    2023

  12. [20]

    Openshape: Scaling up 3d shape representation towards open-world understanding

    Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. Advances in neural information processing systems, 36: 0 44860--44879, 2023

    2023

  13. [21]

    Marching cubes: A high resolution 3d surface construction algorithm

    William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Seminal graphics: pioneering efforts that shaped the field, pp.\ 347--353. 1998

    1998

  14. [22]

    Vehiclesdf: A 3d generative model for constrained engineering design via surrogate modeling

    Hayata Morita, Kohei Shintani, Chenyang Yuan, and Frank Permenter. Vehiclesdf: A 3d generative model for constrained engineering design via surrogate modeling. In NeurIPS 2024 Workshop on Data-driven and Differentiable Simulations, Surrogates, and Solvers, 2024

    2024

  15. [23]

    Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision

    Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 3504--3515, 2020

    2020

  16. [24]

    Deepsdf: Learning continuous signed distance functions for shape representation

    Jeong Joon Park, Peter Florence, and et al. Deepsdf: Learning continuous signed distance functions for shape representation. CVPR, 2019

    2019

  17. [25]

    Pointnet: Deep learning on point sets for 3d classification and segmentation

    Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 652--660, 2017

    2017

  18. [26]

    Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan

    Aditya Sanghi, Hang Chu, Joseph G. Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 18603--1861...

    2022

  19. [27]

    Sketchgraphs: A large-scale dataset for modeling relational geometry in cad

    Ari Seff and et al. Sketchgraphs: A large-scale dataset for modeling relational geometry in cad. In ECCV, 2020

    2020

  20. [28]

    Blendednet: A blended wing body aircraft dataset and surrogate model for aerodynamic predictions

    Nicholas Sung, Steven Spreizer, Mohamed Elrefaie, Kaira Samuel, Matthew C Jones, and Faez Ahmed. Blendednet: A blended wing body aircraft dataset and surrogate model for aerodynamic predictions. arXiv preprint arXiv:2509.07209, 2025

    2025

  21. [29]

    Regression shrinkage and selection via the lasso

    Robert Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B: Statistical Methodology, 58 0 (1): 0 267--288, 1996

    1996

  22. [30]

    Lion: Latent point diffusion models for 3d shape generation

    Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. Advances in Neural Information Processing Systems, 35: 0 10021--10039, 2022

    2022

  23. [31]

    Deepcad: A deep generative network for computer-aided design models

    Chi Wang, Yewen Tian, and et al. Deepcad: A deep generative network for computer-aided design models. In CVPR, 2022

    2022

  24. [32]

    Deep network interpolation for continuous imagery effect transition

    Xintao Wang, Ke Yu, Chao Dong, Xiaoou Tang, and Chen Change Loy. Deep network interpolation for continuous imagery effect transition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 1692--1701, 2019

    2019

  25. [33]

    Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time

    Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing infe...

    2022

  26. [34]

    Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling

    Jiajun Wu and et al. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. In NeurIPS, 2016

    2016

  27. [35]

    Structured 3d latents for scalable and versatile 3d generation

    Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pp.\ 21469--21480, 2025

    2025

  28. [36]

    Learning semantic deformation flows with 3d convolutional networks

    Mehmet Ersin Yumer and Niloy Mitra. Learning semantic deformation flows with 3d convolutional networks. In ECCV, 2016

    2016

  29. [37]

    Lion: Latent point diffusion models for 3d shape generation

    Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems, 2022

    2022

  30. [38]

    Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation

    Zibo Zhao, Wen Liu, Xin Chen, Xianfang Zeng, Rui Wang, Pei Cheng, Bin Fu, Tao Chen, Gang Yu, and Shenghua Gao. Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation. In Advances in Neural Information Processing Systems, 2023

    2023

  31. [39]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  32. [40]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  33. [41]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  34. [42]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.