Pith. sign in

REVIEW 3 major objections 4 minor 35 references

SketchDNN: Joint Continuous-Discrete Diffusion for CAD Sketch Generation

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read SketchDNN, a Gaussian-Softmax diffusion model, generates CAD sketches as sets of primitives and reports FID 7.80 and NLL 81.33 on SketchGraphs, improvements over previous autoregressive baselines.

desk verdict Empirically strong CAD sketch diffusion with a genuinely new continuous-discrete relaxation, but the reverse-process derivation has an algebra error that makes the theory wrong as written. read the letter →

arxiv 2507.11579 v2 pith:EISSDYVL submitted 2025-07-15 cs.CV cs.LG

classification cs.CVcs.LG
keywords CADsketchgenerationdiffusionmodelsdiscreteGaussian-SoftmaxdistributionpermutationequivarianceGraphsprobabilisticgenerativemodelingset
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

SketchDNN is a generative model that produces the geometric primitives of CAD sketches—lines, circles, arcs, and points—as an unordered set instead of a token sequence. Its central claim is that a joint continuous-discrete diffusion process, Gaussian-Softmax diffusion, can model the discrete primitive classes and their continuous coordinates together, and that this beats the previous autoregressive and categorical-diffusion approaches. On the SketchGraphs benchmark the model reports FID (a distribution-distance measure on rendered images) 7.80 versus 16.04 for Vitruvion and NLL 81.33 bits versus 84.8, which the authors take as state-of-the-art evidence for diffusion over autoregressive generation in this domain. The paper also argues, from ablations, that the largest share of the gain comes from superposition—representing each primitive as a mixture of all types—rather than from permutation-invariant denoising alone.

What carries the argument

The central object is the Gaussian-Softmax distribution, the distribution of a Gaussian random vector after the softmax map pushes it onto the probability simplex; mathematically it is a variant of the logistic-normal distribution. It is used twice: in the forward process, where a clean one-hot label is noised as $y_t = \mathrm{softmax}(\sqrt{\alpha_t} \log y_0 + \sqrt{1-\alpha_t} \epsilon)$, and in the reverse process, where the model interpolates between logits of the noisy label and the predicted clean label with a variance-schedule-augmented standard deviation. A transformer without positional encodings plays the role of a permutation-equivariant denoiser, and the continuous parameters of each primitive are diffused with the standard Gaussian process in the same network. The variance schedule augmentation is a separate piece of machinery: it rescales the schedule so that the argmax of the noisy label decays at a controllable rate, which the authors show is not true of the raw cosine schedule.

What would settle it

Sample pairs $(y_0, y_t)$ from the cumulative forward transition, then estimate the conditional mean and variance of $y_{t-1}$ given $(y_t, y_0)$ by Monte Carlo at several timesteps and compare them with the paper's posterior formula. If the measured conditional variance tracks the forward marginal variance $1-\alpha_{t-1}$ rather than the claimed $\sigma_{t-1}^2$, the sampler is not drawing from the posterior the derivation announces; a retrain with the exact conditional variance would then show whether the reported FID gain survives.

Watch

Extended reading notes

Core claim

The paper's contribution is a diffusion model that works directly on CAD sketches and treats each primitive as a composite vector: a construction flag, a class label relaxed to a probability vector, and parameters for every primitive type (line, circle, arc, point). The key behavioral claim is that Gaussian-Softmax diffusion, which adds Gaussian noise to logits and projects them onto the probability simplex with a softmax, gives blended class labels and thereby destroys label information gradually instead of in one step like multinomial diffusion. With a transformer denoiser that omits positional encodings, the whole generation process is permutation equivariant: reordering primitives in the input permutes the output, so ordering never biases the learned distribution. The authors state that this joint continuous-discrete diffusion reduces FID from 16.04 to 7.80 and NLL from 84.8 to 81.33 on SketchGraphs, and their ablations attribute most of the improvement to superposition rather than permutation invariance.

Load-bearing premise

The paper's load-bearing premise is that the reverse step for discrete labels is exactly Gaussian-Softmax with variance $((1-\alpha_t)(1-\alpha_{t-1}))/(1-\alpha_t)$, since every sampling step is built on that closed-form posterior.

Editorial extensions

If this is right

  • If correct, CAD sketch generation can be done as unconditional set generation: one denoising pass produces all primitives without any ordering bias.
  • The reported margins over Vitruvion and SketchGen imply that diffusion with superposition is a stronger paradigm than autoregressive token generation for parametric sketches.
  • The ablation contrast with categorical diffusion suggests that blended class labels, not diffusion per se, are what make the method work on discrete structural labels.
  • Because constraints are excluded from the model, the same framework can be extended to generate the full sketch graph (primitives plus constraints) rather than geometry alone.
  • The variance-schedule augmentation provides a way to control class-label noise in any softmax-projected diffusion model, not just CAD.

Reading between the lines

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

  • Beyond the paper, Gaussian-Softmax diffusion is a natural fit for other set-structured categorical data, such as molecular graphs or UI layouts, where permutation invariance and heterogeneous node types matter in the same way.
  • The paper's schedule augmentation leans on a Gumbel approximation for the argmax of the noisy label; how well the augmented schedule holds for exact Gaussian noise is a question a reader could probe by measuring $p(c_t = c_0)$ directly.
  • A further inference is that latent diffusion underperformed partly because the VAE bottleneck discards geometric detail; a non-latent continuous-only diffusion baseline would isolate whether the discrete modeling or the data-space choice drives the gain.
  • The model currently generates primitives unconditionally; conditioning on a constraint graph or a user-specified subset of primitives is the most immediate next step implied by the failure-case discussion.
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 / 4 minor

Summary. The paper proposes SketchDNN, a generative diffusion model for CAD sketches that jointly handles continuous primitive parameters and discrete class labels through a proposed Gaussian-Softmax diffusion process, and uses a permutation-equivariant transformer denoiser. The method is evaluated on the SketchGraphs dataset, where it reports FID of 7.80 versus 16.04 for the previous state-of-the-art autoregressive model Vitruvion, and NLL of 81.33 versus 84.8, with ablations showing that the Gaussian-Softmax superposition and permutation invariance each contribute to the improvement. The authors also introduce a variance-schedule augmentation heuristic, validated with Monte Carlo simulations, to make the discrete diffusion more gradual in class-label space.

Significance. If the empirical results are reproducible, the contribution is substantial: SketchDNN is the first data-space diffusion model for parametric CAD sketches and it claims large, practically meaningful improvements in both sample fidelity and diversity over strong autoregressive baselines. The idea of performing diffusion on the probability simplex via a softmax transformation of Gaussian-perturbed logits is interesting and could find applications beyond CAD. The paper also contributes a permutation-equivariant denoising architecture and a careful variance-schedule augmentation procedure with Monte Carlo validation. However, the theoretical derivation of the reverse process contains a serious algebraic error that undermines the paper's central claim that its reverse transition is derived from the true posterior, rather than being a heuristic like those in prior simplex-diffusion work. The empirical results may still hold, but the current manuscript does not support its principled-diffusion claim as written.

major comments (3)
  1. [Section 3.3.2, Eq. (7), and Appendix A.3] The posterior variance of the Gaussian-Softmax forward chain is algebraically incorrect. For the forward process with cumulative noise variances sigma_t^2 = 1-alpha_t, the correct posterior variance from completing the square is (1-alpha_t)(1-alpha_{t-1})/(1-alpha_t alpha_{t-1}), not (1-alpha_t)(1-alpha_{t-1})/(1-alpha_t) = 1-alpha_{t-1}. The derivation in Appendix A.3 erroneously simplifies (1/sigma_{t-1}^2 + alpha_t/sigma_t^2)^{-1} to (1-alpha_t)(1-alpha_{t-1})/(1-alpha_t) by dropping the alpha_t sigma_{t-1}^2 term in the denominator. Consequently, the reverse sampling procedure in Eq. (7) does not sample from the posterior of the stated forward process. This is load-bearing because the paper explicitly distinguishes its approach from prior simplex-diffusion methods on the grounds that it derives the reverse transition from the posterior, and the same error appears in the posterior mean. The authors must either correct the posterior derivation and the resulting sampling update or explicitly state that the reverse process is an approximation.
  2. [Section 3.1, Eq. (3)] The continuous-diffusion review contains the same variance error: the expression sqrt((1-alpha_t)(1-alpha_{t-1})/(1-alpha_t)) simplifies to sqrt(1-alpha_{t-1}), which is the forward marginal standard deviation, not the posterior standard deviation of the reverse step. In addition, the notation overloads alpha_t: Eq. (1) uses alpha_t as a per-step noise scale, while Eq. (2) defines alpha_t as the cumulative product, and Eq. (3) then uses alpha_t and alpha_{t-1} as if they were cumulative. This ambiguity propagates to Appendix A.3, where the forward transition p(x_t|x_{t-1}) is written with coefficient sqrt(alpha_t) and variance 1-alpha_t, but the marginal p(x_{t-1}|x_0) is simultaneously treated as having cumulative variance 1-alpha_{t-1}. The paper should adopt distinct notation for per-step and cumulative noise scales (e.g., alpha_t and bar-alpha_t) and re-derive the posterior accordingly.
  3. [Section 5.1 and Tables 1-2] The paper does not specify whether the Vitruvion and SketchGen results in Tables 1 and 2 were obtained by re-evaluating the released baseline models on the same preprocessed test set (the 70K CAD sketches described in Section 5.1) or taken verbatim from the original publications. This matters because the preprocessing used here differs from the original works in several ways: sketches with fewer than eight primitives are removed, duplicates are deduplicated after 8-bit quantization, and coordinates are renormalized to a unit square. If the baseline numbers come from different test sets or preprocessing pipelines, the reported FID and NLL reductions may not reflect a head-to-head comparison. The authors should state the evaluation protocol explicitly and, if the numbers are from the original papers, rerun the baselines on the same test set or clearly qualify the comparison.
minor comments (4)
  1. [Section 3.3 and Conclusion] The manuscript calls the Gaussian-Softmax distribution 'a novel variant of the Logistic-Normal distribution,' but the Gaussian-Softmax distribution is exactly the logistic-normal distribution introduced by Aitchison and Shen (1980); it would be more accurate to describe it as a reparameterization or special case rather than a novel variant.
  2. [Section 6.1] The sentence 'the ELBO, which satisfies the inequality ELBO >= NLL' is confusing: the ELBO lower-bounds the log-likelihood, so the corresponding variational bound on NLL is an upper bound, not a lower bound. The authors should rephrase to avoid implying that the reported NLL values are lower bounds when they are actually variational upper bounds.
  3. [Appendix A.3] The derivation uses sigma_{t-1}^2 both for the marginal variance 1-alpha_{t-1} and for the posterior variance in the same set of equations; using distinct symbols (e.g., s_t^2 for marginal and sigma_{t-1}^2 for posterior) would greatly improve clarity and prevent the kind of cancellation error noted in the major comments.
  4. [Figure 2] The left and right panels of Figure 2 are referenced in the caption and text as showing the raw and augmented schedules, but the panel labels themselves are not shown; adding small 'raw' and 'augmented' labels to the panels would make the figure easier to read.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's central FID/NLL claims are measured on a held-out test set against external baselines; the reverse-process algebra error is a correctness issue, not a self-referential reduction.

full rationale

The central claims are empirical: FID 7.80 versus 16.04 and NLL 81.33 versus 84.80, evaluated on a held-out 70K-sketch test set against the external baselines Vitruvion and SketchGen, plus internal ablations that share architecture and training details. No reported number comes from fitting a parameter to the test set; k = 0.99, lambda = 16, and the augmented variance schedule are fixed constants or calibration heuristics, not test-set-fitted parameters. I checked the derivation chain for self-reference. The cumulative forward transition (Eq. 6) is derived from the per-step transition (Eq. 5) in Appendix A.1 by Gaussian composition and softmax shift-invariance; the reverse posterior in Appendix A.3 is derived from the forward densities by completing the square, so the sampler is not, by construction, identical to the training objective. There are no load-bearing self-citations: the authors cite no prior work of their own, and the invoked results (DDPM posterior form, Logistic-Normal density, transformer permutation equivariance) are external. One legitimate concern is that the posterior-variance algebra in Section 3.3.2 and Appendix A.3 is incorrect: with sigma_t^2 = 1 - a_t, the precision sum is 1/sigma_{t-1}^2 + a_t/sigma_t^2 = (1 - a_t a_{t-1}) / ((1 - a_t)(1 - a_{t-1})), whose reciprocal is not the paper's (1 - a_t)(1 - a_{t-1})/(1 - a_t), so the stated sigma_{t-1}^2 = 1 - a_{t-1} does not follow from the stated forward chain. This is a mathematical-correctness and reproducibility issue, not a circularity: the sampler is not forced to match the data by definition, and the empirical FID/NLL results stand or fall independently of this algebra. No pattern from the circularity taxonomy is exhibited.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim depends on the Gaussian-Softmax reverse-process derivation (Appendix A.3), the variance schedule augmentation (Eq. 8), and the composite representation; the free parameters k, lambda, and n are hand-chosen but not fitted to the test set.

free parameters (3)
  • k = 0.99
    Smoothing constant for one-hot class labels in the cumulative forward transition (Eq. 6), chosen to avoid log(0) singularities. Not fitted to data but a hand-set hyperparameter.
  • lambda = 16
    Weight for the MSE loss for timesteps t <= 150 in the reconstruction loss (Section 5.2), chosen by the authors based on preliminary experiments.
  • max_primitives n = 16
    Maximum number of primitives in a sketch, set due to resource constraints and used to define the fixed-size representation (Section 5.1). This is a data-processing choice that determines the model's output dimension.
assumptions (4)
  • domain assumption The forward process in logit space is a linear Gaussian diffusion with per-step coefficient sqrt(alpha_t) and variance 1-alpha_t, and the cumulative transition sums independent Gaussian noises to variance 1 - product(alpha_i).
    This is the foundation for deriving the cumulative transition and posterior; it assumes the log-softmax projection preserves the Gaussian linear structure.
  • ad hoc to paper Approximation of argmax of Gaussian vector by argmax of Gumbel vector to calibrate the variance schedule (Appendix A.4).
    The variance schedule augmentation is derived by assuming the argmax of a Gaussian can be modeled by a Gumbel reparameterization; this is an approximation, not an exact equivalence.
  • ad hoc to paper The Gaussian-Softmax distribution's posterior after a Markov step is again Gaussian-Softmax with the derived mean and variance.
    This is the core theoretical claim of the paper; the derivation in Appendix A.3 uses particular transition densities and a change of variables, but the soundness of this result is questionable as discussed in red flags.
  • domain assumption CAD sketches can be represented as a set of up to 16 primitives, each modeled as a superposition of all primitive types, and that this composite representation is sufficient for the generative task.
    The paper defines the sketch representation in Section 2 and uses it throughout; the adequacy of this representation is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SketchDNN: Joint Continuous-Discrete Diffusion for CAD Sketch Generation." pith.science (2026). https://pith.science/paper/EISSDYVL

@misc{pith2026250711579,
  author       = {Pith},
  title        = {Pith review of: SketchDNN: Joint Continuous-Discrete Diffusion for CAD Sketch Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EISSDYVL}},
  note         = {Machine review of arXiv:2507.11579}
}
read the original abstract

We present SketchDNN, a generative model for synthesizing CAD sketches that jointly models both continuous parameters and discrete class labels through a unified continuous-discrete diffusion process. Our core innovation is Gaussian-Softmax diffusion, where logits perturbed with Gaussian noise are projected onto the probability simplex via a softmax transformation, facilitating blended class labels for discrete variables. This formulation addresses 2 key challenges, namely, the heterogeneity of primitive parameterizations and the permutation invariance of primitives in CAD sketches. Our approach significantly improves generation quality, reducing Fr\'echet Inception Distance (FID) from 16.04 to 7.80 and negative log-likelihood (NLL) from 84.8 to 81.33, establishing a new state-of-the-art in CAD sketch generation on the SketchGraphs dataset.

Figures

Figures reproduced from arXiv: 2507.11579 by the authors.

Figure 1
Figure 1. CAD sketches generated by our diffusion model, showcasing its ability to produce diverse, high-fidelity designs. Geometric primitives—such as circles, arcs, lines, and points—are randomly colored to differentiate separate primitives, and primitives tagged as construction aids are represented with dashed lines. Abstract We present SketchDNN, a generative model for synthesizing CAD sketches that jointly models both co… view at source ↗
Figure 2
Figure 2. Left: The orange curve represents the raw cosine variance schedule at, while the blue curve depicts the probability that the class label remains unchanged. Right: The orange curve shows the augmented variance schedule bt, as defined in Equation (8), and the blue curve again represents the probability of the class label not switching. The probabilities were calculated using Monte Carlo estimation over 100,000 samples… view at source ↗
Figure 3
Figure 3. The generation pipeline of SketchDNN. Starting from a pure noise seed XT , the denoiser network iteratively refines the sample by interpolating between the noisy input and the model’s prediction of the clean output. The final generated sketch, X0, is obtained after T successive denoising steps. The interpolation formulae are given by Equation 7 for discrete variables and Equation 3 for continuous variables. 5. Exper… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Comparison of CAD sketches from the SketchGraphs dataset (top), generations from SketchDNN (middle), and generations from Vitruvion (bottom) [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Visualizations of 3 primary failure cases. Left: Gen￾erated CAD sketch with no discernible pattern. Middle: Large gaps between primitives. Right: Extraneous primitives, zoom of the bottom left portion of a triangle where the bottom leg has an overlapping redundant line…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 28 canonical work pages

  1. [1]

    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 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    and Shen, S

    Aitchison, J. and Shen, S. M. Logistic-normal distributions: Some properties and uses. Biometrika, 67 0 (2): 0 261--272, 1980. ISSN 00063444, 14643510

  3. [3]

    D., Ho, J., Tarlow, D., and van den Berg, R

    Austin, J., Johnson, D. D., Ho, J., Tarlow, D., and van den Berg, R. Structured denoising diffusion models in discrete state-spaces. In Beygelzimer, A., Dauphin, Y., Liang, P., and Vaughan, J. W. (eds.), Advances in Neural Information Processing Systems, 2021

  4. [4]

    Bond-Taylor, S., Hessey, P., Sasaki, H., Breckon, T., and Willcocks, C. G. Unleashing transformers: Parallel token prediction with discrete absorbing diffusion for fast high-resolution image generation from vector-quantized codes. In European Conference on Computer Vision, 2021

  5. [5]

    Play: Parametrically conditioned layout generation using latent diffusion

    Cheng, C.-Y., Huang, F., Li, G., and Li, Y. Play: Parametrically conditioned layout generation using latent diffusion. In International Conference on Machine Learning, 2023

  6. [6]

    L., Ollion, C., and Moulines, E

    Cohen, M., Quispe, G., Corff, S. L., Ollion, C., and Moulines, E. Diffusion bridges vector quantized variational autoencoders. In Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., and Sabato, S. (eds.), Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, pp.\ 4141--4156...

  7. [7]

    Computer-aided design as language

    Ganin, Y., Bartunov, S., Li, Y., Keller, E., and Saliceti, S. Computer-aided design as language. In Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., and Vaughan, J. W. (eds.), Advances in Neural Information Processing Systems, volume 34, pp.\ 5885--5897. Curran Associates, Inc., 2021

  8. [8]

    Ssd-lm: Semi-autoregressive simplex-based diffusion language model for text generation and modular control

    Han, X., Kumar, S., and Tsvetkov, Y. Ssd-lm: Semi-autoregressive simplex-based diffusion language model for text generation and modular control. In Annual Meeting of the Association for Computational Linguistics, 2022

Show all 35 references
  1. [9]

    Denoising diffusion probabilistic models

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. In Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., and Lin, H. (eds.), Advances in Neural Information Processing Systems, volume 33, pp.\ 6840--6851. Curran Associates, Inc., 2020

  2. [10]

    Argmax flows and multinomial diffusion: Learning categorical distributions

    Hoogeboom, E., Nielsen, D., Jaini, P., Forr \'e , P., and Welling, M. Argmax flows and multinomial diffusion: Learning categorical distributions. In Beygelzimer, A., Dauphin, Y., Liang, P., and Vaughan, J. W. (eds.), Advances in Neural Information Processing Systems, 2021

  3. [11]

    A., Bastings, J., Poole, B., van den Berg, R., and Salimans, T

    Hoogeboom, E., Gritsenko, A. A., Bastings, J., Poole, B., van den Berg, R., and Salimans, T. Autoregressive diffusion models. In International Conference on Learning Representations, 2022

  4. [12]

    Huijben, I. A. M., Kool, W., Paulus, M. B., and van Sloun, R. J. G. A review of the gumbel-max trick and its extensions for discrete stochasticity in machine learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45: 0 1353--1371, 2021

  5. [13]

    Categorical reparameterization with gumbel-softmax

    Jang, E., Gu, S., and Poole, B. Categorical reparameterization with gumbel-softmax. In International Conference on Learning Representations, 2017

  6. [14]

    TESS : Text-to-text self-conditioned simplex diffusion

    Karimi Mahabadi, R., Ivison, H., Tae, J., Henderson, J., Beltagy, I., Peters, M., and Cohan, A. TESS : Text-to-text self-conditioned simplex diffusion. In Graham, Y. and Purver, M. (eds.), Proceedings of the 18th Conference of the European Chapter of the Association for Comput...

  7. [15]

    Crafting papers on machine learning

    Langley, P. Crafting papers on machine learning. In Langley, P. (ed.), Proceedings of the 17th International Conference on Machine Learning (ICML 2000), pp.\ 1207--1216, Stanford, CA, 2000. Morgan Kaufmann

  8. [16]

    Dlt: Conditioned layout generation with joint discrete-continuous diffusion layout transformer

    Levi, E., Brosh, E., Mykhailych, M., and Perez, M. Dlt: Conditioned layout generation with joint discrete-continuous diffusion layout transformer. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pp.\ 2106--2115, 2023

  9. [17]

    Discrete diffusion modeling by estimating the ratios of the data distribution

    Lou, A., Meng, C., and Ermon, S. Discrete diffusion modeling by estimating the ratios of the data distribution. In Salakhutdinov, R., Kolter, Z., Heller, K., Weller, A., Oliver, N., Scarlett, J., and Berkenkamp, F. (eds.), Proceedings of the 41st International Conference on Ma...

  10. [18]

    J., Mnih, A., and Teh, Y

    Maddison, C. J., Mnih, A., and Teh, Y. W. The concrete distribution: A continuous relaxation of discrete random variables. In International Conference on Learning Representations, 2017

  11. [19]

    Nichol, A. Q. and Dhariwal, P. Improved denoising diffusion probabilistic models. In Meila, M. and Zhang, T. (eds.), Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pp.\ 8162--8171. PMLR, 18--24 Jul 2021

  12. [20]

    R., Bhat, S

    Para, W. R., Bhat, S. F., Guerrero, P., Kelly, T., Mitra, N., Guibas, L., and Wonka, P. Sketchgen: generating constrained cad sketches. In Proceedings of the 35th International Conference on Neural Information Processing Systems, NIPS '21, Red Hook, NY, USA, 2024. Curran Assoc...

  13. [21]

    Peebles, W. S. and Xie, S. Scalable diffusion models with transformers. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pp.\ 4172--4182, 2022

  14. [22]

    Potapczynski, A., Loaiza-Ganem, G., and Cunningham, J. P. Invertible gaussian reparameterization: revisiting the gumbel-softmax. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS '20, Red Hook, NY, USA, 2020. Curran Associates I...

  15. [23]

    Seff, A., Ovadia, Y., Zhou, W., and Adams, R. P. Sketchgraphs: A large-scale dataset for modeling relational geometry in computer-aided design. ArXiv, abs/2007.08506, 2020

  16. [24]

    Seff, A., Zhou, W., Richardson, N., and Adams, R. P. Vitruvion: A generative model of parametric CAD sketches. In International Conference on Learning Representations, 2022

  17. [25]

    A., Hosseini, S., and Furukawa, Y

    Shabani, M. A., Hosseini, S., and Furukawa, Y. Housediffusion: Vector floorplan generation via a diffusion model with discrete and continuous denoising. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 5466--5475, 2022

  18. [26]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., and Ganguli, S. Deep unsupervised learning using nonequilibrium thermodynamics. In Bach, F. and Blei, D. (eds.), Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings of Machine Learni...

  19. [27]

    N., Kaiser, L., and Polosukhin, I

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., and Polosukhin, I. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, pp.\ 6000–6010, Red Hook, NY, USA, 201...

  20. [28]

    Vq-cad: Computer-aided design model generation with vector quantized diffusion

    Wang, H., Zhao, M., Wang, Y., Quan, W., and Yan, D.-M. Vq-cad: Computer-aided design model generation with vector quantized diffusion. Comput. Aided Geom. Des., 111: 0 102327, 2024

  21. [29]

    Willis, K. D. D., Jayaraman, P. K., Lambourne, J., Chu, H., and Pu, Y. Engineering sketch generation for computer-aided design. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pp.\ 2105--2114, 2021

  22. [30]

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

    Wu, R., Xiao, C., and Zheng, C. Deepcad: A deep generative network for computer-aided design models. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pp.\ 6752--6762, 2021

  23. [31]

    H., Katz, M., Jayaraman, P

    Wu, S., Khasahmadi, A. H., Katz, M., Jayaraman, P. K., Pu, Y., Willis, K., and Liu, B. Cadvlm: Bridging language and vision in the generation of parametric cad sketches. In Leonardis, A., Ricci, E., Roth, S., Russakovsky, O., Sattler, T., and Varol, G. (eds.), Computer Vision ...

  24. [32]

    Permutation equivariance of transformers and its applications

    Xu, H., Xiang, L., Ye, H., Yao, D., Chu, P., and Li, B. Permutation equivariance of transformers and its applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 5987--5996, June 2024 a

  25. [33]

    Xu, X., Willis, K. D. D., Lambourne, J., Cheng, C.-Y., Jayaraman, P. K., and Furukawa, Y. Skexgen: Autoregressive generation of cad construction sequences with disentangled codebooks. ArXiv, abs/2207.04632, 2022

  26. [34]

    K., Wang, Z., Willis, K

    Xu, X., Lambourne, J., Jayaraman, P. K., Wang, Z., Willis, K. D. D., and Furukawa, Y. Brepgen: A b-rep generative diffusion model with structured latent geometry. ACM Transactions on Graphics (TOG), 43: 0 1 -- 14, 2024 b

  27. [35]

    Yu, P., Xie, S., Ma, X., Jia, B., Pang, B., Gao, R., Zhu, Y., Zhu, S.-C., and Wu, Y. N. Latent diffusion energy-based model for interpretable text modelling. In Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., and Sabato, S. (eds.), Proceedings of the 39th Intern...

Pith tools

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