Pith. sign in

REVIEW 4 major objections 4 minor 52 references

PyFRep: Shape Modeling with Differentiable Function Representation

T0 review · 4 major / 4 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read PyFRep makes Function Representation shapes differentiable through automatic differentiation, giving derivatives with respect to space or shape parameters, and demonstrates them on curvature, signed distance, and parametric fitting.

desk verdict PyFRep is a genuinely useful open-source FRep toolkit with correct autodiff, but the paper's experimental validation is mostly qualitative and the fitting experiment doesn't isolate gradients from the evolutionary warm start. read the letter →

arxiv 2504.14216 v1 pith:WTOIIAKT submitted 2025-04-19 cs.GR

classification cs.GR
keywords ShapemodelingFunctionRepresentationImplicitsurfaceDifferentiableAutomaticdifferentiationSigneddistanceParametricmodelfittingCurvatureestimation
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

PyFRep is a framework for differentiable geometric modeling built on Function Representation, where a solid is defined by a scalar function $f:\mathbb{R}^3\to\mathbb{R}$ whose zero level-set is the surface. The paper's claim is that by writing this function as a Python program and tracing it with an automatic-differentiation library, the same shape definition yields derivatives with respect to spatial coordinates or shape parameters. A sympathetic reader would care because these derivatives turn shape programs into interrogated and optimizable objects: they support curvature computation, distance-function approximation, and fitting parametric models to point clouds. The framework is released as open source, so the claim is testable and extensible.

What carries the argument

The central object is the FRep scalar function $f:\mathbb{R}^3\to\mathbb{R}$, whose zero level-set is the surface, written in a subset of Python and evaluated under a tensor-based automatic-differentiation library that traces the program; in the current implementation this library is PyTorch [39]. Reverse-mode automatic differentiation then produces first- and second-order derivatives of $f$ with respect to the input coordinates or to parameters stored in the computational graph. From the gradient and divergence alone the framework computes mean curvature via $H=-\tfrac12\operatorname{div}(\nabla f/|\nabla f|)$, Gaussian curvature via the adjoint-Hessian formula, the Laplacian, normalization schemes $\omega_1$ and $\delta_1$, and the loss gradients used for fitting. The key design choice is that all primitives and operations are pure Python, so the derivative computation is decoupled from the shape definition and any automatic-differentiation backend can be swapped in.

What would settle it

Take a parametric union of two spheres created with a min operation and sample the input point cloud densely along the intersection circle, where the FRep function is not differentiable; if SGD driven by the library's subgradients cannot recover the correct radii and positions or stalls near the crease, the framework's claim that non-differentiability did not pose any problem would be refuted for a common modeling operation.

Watch

Extended reading notes

Core claim

The central claim is that an FRep shape defined by a program $f:\mathbb{R}^3\to\mathbb{R}$ that is differentiable almost everywhere can be made fully differentiable: reverse-mode automatic differentiation over the program's computational graph supplies first- and second-order derivatives of $f$ with respect to the point $x$, for shape interrogation, and with respect to parameters $p$, for shape optimization. The paper demonstrates that these derivatives are sufficient to compute mean, Gaussian, and principal curvatures at surface vertices via $H=-\tfrac12\operatorname{div}(\nabla f/|\nabla f|)$ and the adjoint-Hessian expression for $K$; to estimate signed distance by normalizing $f$ with Rvachev or Taubin schemes or by training a neural network $h$ with the loss $\mathbb{E}[(|\nabla d|-1)^2]$ while preserving the zero level-set through $d=\operatorname{sign}(f)h$; and to fit a parametric microstructure to a $40{,}000$-point cloud by minimizing $N^{-1}\sum_i f(x_i;p)^2$ with regularized evolution plus SGD. The claim is that all of this works without leaving the FRep programming model, and that the pure-Python primitive layer can be swapped across automatic-differentiation backends.

Load-bearing premise

The framework assumes that the shape-defining function is differentiable almost everywhere and that the automatic-differentiation library's subgradients at non-differentiable points, such as the creases created by min/max Boolean operations, behave well enough for optimization; the paper supports this only by saying that its experiments encountered no problems.

Editorial extensions

If this is right

  • Curvature and other differential surface quantities can be computed directly from the defining program, with no mesh fitting or polynomial approximation step, for any FRep shape.
  • Signed distance functions for implicitly defined surfaces can be produced either by closed-form normalization or by training a neural network with the eikonal loss, using automatic differentiation for both.
  • Parametric FRep models can be fitted to raw point clouds by minimizing the squared function values, and the regularized-evolution-plus-SGD schedule escapes local minima that plain SGD gets stuck in.
  • Because the shape program is pure Python and derivatives are obtained by tracing, the same modeling code can be reused with different automatic-differentiation backends, not only the one used here.
  • Shape parameters can be optimized against any differentiable objective, not just point-cloud distance, opening the framework to downstream geometric pipelines.

Reading between the lines

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

  • Editorial inference: the same machinery could support inverse procedural modeling at scale, recovering CSG or FRep parameter values of a scanned object directly from a point cloud or depth image, which is exactly the optimization the fitting example demonstrates on a microstructure.
  • Editorial inference: since curvature and the Laplacian come from automatic differentiation of the shape program, the framework could serve as a differentiable regularizer for shape fairness or for penalizing high curvature during optimization, a use the paper does not explore.
  • Editorial inference: a natural stress test would be a benchmark across a family of shapes with creases, comparing the chosen subgradients against smooth $C^1$ approximations of min/max to see where the almost-everywhere differentiability assumption actually matters.
  • Editorial inference: the backend-agnostic pure-Python core suggests a compiler or JIT target could accelerate the traced programs, and the neural SDF approach hints at a hybrid where the network itself is optimized in the same differentiable loop as the shape parameters.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. This paper introduces PyFRep, a Python/FRep framework that represents solids as scalar functions and uses PyTorch automatic differentiation to obtain derivatives with respect to spatial coordinates or shape parameters. The authors describe the architecture, primitives, operations (including CSG, SDF, periodic functions), and derivative computations, and demonstrate four applications: mean/Gaussian/principal curvature estimation, Rvachev/Taubin normalization, signed distance function estimation via a neural network ansatz, and parametric shape fitting to point clouds using regularized evolution followed by SGD. The code is released as open source.

Significance. PyFRep addresses a useful niche: combining FRep constructive modeling with automatic differentiation in a single open-source package. Its clean separation of pure-Python primitives from autodiff backends and its use of exact derivative formulas rather than mesh-based approximations are genuine strengths. If the empirical claims are substantiated, the framework could be a convenient platform for shape interrogation and inverse modeling. However, the paper's evidence is mostly qualitative: curvature errors are judged visually, SDF results are compared by contour plots only, and the fitting experiment does not quantify the contribution of gradients versus the evolutionary warm start. These gaps concern the demonstrated applications that constitute the paper's central contribution, not merely the presentation.

major comments (4)
  1. [§5.4, Eq. (10)] The fitting experiment does not isolate the contribution of differentiable gradients. The pipeline uses 10,000 regularized-evolution iterations to obtain the initial parameters p0 and then 100 SGD steps; the 'SGD alone' baseline is not specified in terms of initialization, learning rate, batch size, or random seed, and the results are reported only as color maps of absolute pointwise error with no summary error, loss curves, or restarts. Since min/max-based CSG can yield zero subgradients with respect to shape parameters over open regions, the reported improvement may come largely from the evolutionary warm start. Please report quantitative errors, an ablated SGD run from the same p0, longer SGD (or loss-vs-iteration curves), and a diagnostic of the fraction of parameters with zero gradient during optimization.
  2. [§4] The claim that the lack of differentiability 'did not pose any problem based on our experiments and in the applications that we considered' is anecdotal and load-bearing for the framework's reliability. For min/max Boolean operations, PyTorch's subgradient can be zero over whole active regions, so shape-parameter optimization can stall; the paper should test this systematically, for example by running several random restarts on a few CSG models with different loss functions and recording convergence or the frequency of zero-gradient steps.
  3. [§5.3, Eq. (9)] The signed-distance estimation is validated only visually ('Compare the filled contour plots with Fig. 7'), with no error metric against a known distance field. Because the method follows the author's prior work [11], the framework contribution would be better supported by a quantitative comparison (e.g., mean absolute error to the exact SDF for a sphere or ellipse, or convergence of the loss (9)).
  4. [§5.1, Eqs. (3)–(4)] Curvature evaluation is verified only by stating that mean-curvature values on the Schwarz D surface are 'close to zero' and by a histogram without axis scales. Since curvatures are a headline application, report quantitative errors against analytic values (e.g., H=1, K=1 for a unit sphere) or a mesh-refinement convergence study.
minor comments (4)
  1. [Fig. 5] Add axis labels and units to the histogram and report the mean and standard deviation of the computed mean-curvature values.
  2. [§5.4] For reproducibility, spell out all hyperparameters for regularized evolution and SGD, including population size, tournament sample size, mutation details, learning rate, batch size, number of epochs, and the initialization used for the 'SGD alone' baseline.
  3. [§5.4] Use consistent notation for quantities such as '40 K' and '10 K'; write '40K' and '10K' uniformly.
  4. [References] Reference [11] is an arXiv v1 preprint from 2021; if a peer-reviewed version exists, consider citing it instead.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an engineering/system contribution whose examples validate against independent mathematics and standard losses, with no fitted quantity presented as a prediction.

full rationale

The paper's central claim is that PyFRep provides a differentiable Function Representation framework built on automatic differentiation, and the applications are demonstrations rather than derived predictions. The curvature computation uses the standard formulas from Goldman [16] and is validated against the known result H = 0 on the Schwarz D minimal surface, which is independent external mathematics, not an input of the method. The SDF estimation uses the author's prior method [11], but the paper explicitly states the ansatz sign(f(x))h(x; theta) and the eikonal loss (9), and uses it as a component; the self-citation is real methodological support, not a claim that reduces to itself. The parametric fitting minimizes the standard least-squares loss E(p) = (1/N) sum f(x_i; p)^2, and the reported comparison between regularized evolution + SGD and SGD alone is an experimental evaluation, not a prediction derived from the fitted values. The skeptic's concern that min/max gradients vanish and that the evolutionary warm start carries much of the optimization is a robustness/correctness issue, not circularity: no equation or fitted parameter is equivalently reused as the claimed output. Therefore no circular step can be exhibited and the appropriate finding is no significant circularity.

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

No new physical entities or hand-fitted constants are introduced. The framework is an implementation artifact. The main assumptions are the differentiability of shape functions and the representational power of neural networks for signed distance estimation.

assumptions (4)
  • domain assumption Shape-defining functions f are differentiable almost everywhere.
    The framework and its derivative computations rely on differentiability except at measure-zero sets; the paper states this in Section 4 and relies on subgradients from PyTorch at non-differentiable points.
  • domain assumption The zero level-set of f defines a regular surface with a well-defined unit normal n = ∇f / |∇f| where ∇f ≠ 0.
    Curvature and normalization formulas in Sections 5.1 and 5.2 require a non-vanishing gradient on the surface; this is assumed without explicit safeguards for degenerate points.
  • domain assumption The neural network h(x;θ) can approximate the signed distance function to the zero level-set of f.
    Section 5.3 assumes this representational capability, following the author's prior work [11]; the paper does not independently validate it.
  • standard math The chain rule of differentiation holds for the composed primitive functions.
    Automatic differentiation in Section 3.2 is built on this standard result.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PyFRep: Shape Modeling with Differentiable Function Representation." pith.science (2026). https://pith.science/paper/WTOIIAKT

@misc{pith2026250414216,
  author       = {Pith},
  title        = {Pith review of: PyFRep: Shape Modeling with Differentiable Function Representation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WTOIIAKT}},
  note         = {Machine review of arXiv:2504.14216}
}
read the original abstract

We propose a framework for performing differentiable geometric modeling based on the Function Representation (FRep). The framework is built on top of modern libraries for performing automatic differentiation allowing us to obtain derivatives w.r.t. space or shape parameters. We demonstrate possible applications of this framework: Curvature estimation for shape interrogation, signed distance function computation and approximation and fitting shape parameters of a parametric model to data. Our framework is released as open-source.

Figures

Figures reproduced from arXiv: 2504.14216 by the authors.

Figure 1
Figure 1. Example of a program defining a simple shape (left) and the [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Different shapes obtained by varying the radius of the cylinders [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Computation of ∇f(x). gradient of a model is useful w.r.t. both its spatial parameters or its shape parameters (for shape optimization). For the latter, usually one considers the gradient of a loss function, such as (2), with respect to the shape param￾eters p (the spatial parameters x are fixed). We also provide functions for computing the divergence, the Laplacian (the divergence of the gradient), 9 [PITH_FULL_IM… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Computation of the divergence of a vector field [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Left: Mean curvature on a minimal surface (the Schwarz D mini [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Left: Level-sets (from -0.5 to 0.5) of an ellipse with semi-major [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: shows a chess pawn modeled implicitly, using simple primitives (polynomials) and Boolean operations. The image on the left illustrates the surface {x ∈ R 3 : f(x) = 0} and the filled contour plot of f on a slice. The image on the right shows the filled contour plot of …
Figure 8
Figure 8. Figure 8: Left: Zero level-set of d(x) = sign(f(x))h(x; θe) and filled contour plot on a slice. Right: Filled contour plot on a slice without the surface. parametric model, where p are the parameters controlling the shape of the object. For a given set of parameters p, the corre…
Figure 9
Figure 9. Figure 9: Point-wise absolute error for the fitted model. Left: Regular [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: Reconstructed object from fitting a parametric model. Left: [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 48 canonical work pages

  1. [11]

    Signed distance function computation from an implicit surface

    Pierre-Alain Fayolle. Signed distance function computation from an implicit surface. ArXiv preprint arXiv:2104.08057v1 [cs.GR] , 2021

  2. [1]

    https://openscad.org/

    OpenSCAD. https://openscad.org/

  3. [2]

    Ten- sorFlow: Large-scale machine learning on heterogeneous systems, 2015

    Mart´ ın Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, and et al. Ten- sorFlow: Large-scale machine learning on heterogeneous systems, 2015. Soft- ware available from tensorflow.org

  4. [3]

    Aether: An embedded domain specific sampling language for monte carlo rendering

    Luke Anderson, Tzu-Mao Li, Jaakko Lehtinen, and Fr´ edo Durand. Aether: An embedded domain specific sampling language for monte carlo rendering. ACM Transactions on Graphics (TOG) , 36(4):1–16, 2017. 17

  5. [4]

    Pearlmutter, Alexey Andreyevich Radul, and Jeffrey Mark Siskind

    Atılım G¨ unes Baydin, Barak A. Pearlmutter, Alexey Andreyevich Radul, and Jeffrey Mark Siskind. Automatic differentiation in machine learning: A survey. J. Mach. Learn. Res. , 18(1):5595–5637, January 2017

  6. [5]

    Approximate distance fields with non- vanishing gradients

    Arpan Biswas and Vadim Shapiro. Approximate distance fields with non- vanishing gradients. Graphical Models, 66(3):133–159, 2004

  7. [6]

    Heterogeneous material modeling with distance fields

    Arpan Biswas, Vadim Shapiro, and Igor Tsukanov. Heterogeneous material modeling with distance fields. Computer Aided Geometric Design , 21(3):215– 242, 2004

  8. [7]

    Interactive techniques for implicit mod- eling

    Jules Bloomenthal and Brian Wyvill. Interactive techniques for implicit mod- eling. SIGGRAPH Comput. Graph. , 24(2):109–116, February 1990

Show all 52 references
  1. [8]

    JAX: composable transformations of Python+NumPy programs

    James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, and et al. JAX: composable transformations of Python+NumPy programs. http: //github.com/google/jax, 2018

  2. [9]

    Web-based shape modeling with hyperfun

    Richard Cartwright, Valery Adzhiev, Alexander A Pasko, Yuichiro Goto, and Tosiyasu L Kunii. Web-based shape modeling with hyperfun. IEEE Computer Graphics and Applications, 25(2):60–69, 2005

  3. [10]

    Opt: A domain specific language for non-linear least squares optimization in graphics and imaging

    Zachary DeVito, Michael Mara, Michael Zollh¨ ofer, Gilbert Bernstein, Jonathan Ragan-Kelley, Christian Theobalt, Pat Hanrahan, Matthew Fisher, and Matthias Niessner. Opt: A domain specific language for non-linear least squares optimization in graphics and imaging. ACM Transact...

  4. [12]

    A survey of methods for converting unstructured data to csg models

    Pierre-Alain Fayolle and Markus Friedrich. A survey of methods for converting unstructured data to csg models. Computer-Aided Design, 168:103655, 2024

  5. [13]

    Automation of the Volumetric Models Con- struction, pages 214–238

    Pierre-Alain Fayolle, Alexander Pasko, Elena Kartasheva, Christophe Rosen- berger, and Christian Toinard. Automation of the Volumetric Models Con- struction, pages 214–238. Springer Berlin Heidelberg, Berlin, Heidelberg, 2008

  6. [14]

    Constructive heterogeneous object modeling using signed approxi- mate real distance functions

    Pierre-Alain Fayolle, Alexander Pasko, Benjamin Schmitt, and Nikolay Mirenkov. Constructive heterogeneous object modeling using signed approxi- mate real distance functions. Journal of Computing and Information Science in Engineering, 6(3):221–229, 11 2005

  7. [15]

    Web-based constructive shape modeling using real distance functions

    Pierre-Alain Fayolle, Benjamin Schmitt, Yuichiro Goto, and Alexander Pasko. Web-based constructive shape modeling using real distance functions. IEICE TRANSACTIONS on Information and Systems , 88(5):828–835, 2005

  8. [16]

    Curvature formulas for implicit curves and surfaces

    Ron Goldman. Curvature formulas for implicit curves and surfaces. Computer Aided Geometric Design, 22(7):632–658, 2005. 18

  9. [17]

    Implicit Curves and Surfaces: Mathematics, Data Structures and Al- gorithms

    Abel Gomes, Irina Voiculescu, Joaquim Jorge, Brian Wyvill, and Callum Gal- braith. Implicit Curves and Surfaces: Mathematics, Data Structures and Al- gorithms. Springer Publishing Company, Incorporated, 1st edition, 2009

  10. [18]

    Evaluating Derivatives: Principles and Techniques of Algorithmic Differentiation

    Andreas Griewank and Andrea Walther. Evaluating Derivatives: Principles and Techniques of Algorithmic Differentiation . Society for Industrial and Ap- plied Mathematics, USA, second edition, 2008

  11. [19]

    A bayesian inference framework for procedural material parameter estimation

    Yu Guo, Milos Hasan, Lingqi Yan, and Shuang Zhao. A bayesian inference framework for procedural material parameter estimation. Computer Graphics Forum, 39(7):255–266, 2020

  12. [20]

    X-cad: optimizing cad models with extended finite elements

    Christian Hafner, Christian Schumacher, Espen Knoop, Thomas Auzinger, Bernd Bickel, and Moritz B¨ acher. X-cad: optimizing cad models with extended finite elements. ACM Transactions on Graphics (TOG) , 38(6):1–15, 2019

  13. [21]

    An overview of materials with triply periodic minimal surfaces and related geometry: From biological structures to self-assembled systems

    Lu Han and Shunai Che. An overview of materials with triply periodic minimal surfaces and related geometry: From biological structures to self-assembled systems. Advanced Materials, 30(17):1:705–708, 2018

  14. [22]

    Sphere tracing: A geometric method for the antialiased ray tracing of implicit surfaces

    John C Hart. Sphere tracing: A geometric method for the antialiased ray tracing of implicit surfaces. The Visual Computer , 12(10):527–545, 1996

  15. [23]

    Difftaichi: Differentiable programming for physical simulation

    Yuanming Hu, Luke Anderson, Tzu-Mao Li, Qi Sun, Nathan Carr, Jonathan Ragan-Kelley, and Fr´ edo Durand. Difftaichi: Differentiable programming for physical simulation. ArXiv preprint arXiv:1910.00935, 2019

  16. [24]

    Differentiable rendering: A survey

    Hiroharu Kato, Deniz Beker, Mihai Morariu, Takahiro Ando, Toru Matsuoka, Wadim Kehl, and Adrien Gaidon. Differentiable rendering: A survey. Arxiv prepring arxiv:2006.12057 [cs.CV] , 2020

  17. [25]

    Matt Keeter. LibFive. https://github.com/libfive/libfive

  18. [26]

    Massively parallel rendering of complex closed-form implicit surfaces

    Matt Keeter. Massively parallel rendering of complex closed-form implicit surfaces. ACM Trans. Graph., 39(4), July 2020

  19. [27]

    IceSL: A GPU accelerated CSG modeler and slicer

    Sylvain Lefebvre. IceSL: A GPU accelerated CSG modeler and slicer. In 18th European Forum on Additive Manufacturing (AEFA’13) . AEFA, 2013

  20. [28]

    Differen- tiable monte carlo ray tracing through edge sampling

    Tzu-Mao Li, Miika Aittala, Fr´ edo Durand, and Jaakko Lehtinen. Differen- tiable monte carlo ray tracing through edge sampling. ACM Transactions on Graphics (TOG), 37(6):1–11, 2018

  21. [29]

    Differentiable programming for image processing and deep learning in halide

    Tzu-Mao Li, Micha¨ el Gharbi, Andrew Adams, Fr´ edo Durand, and Jonathan Ragan-Kelley. Differentiable programming for image processing and deep learning in halide. ACM Transactions on Graphics (TOG) , 37(4):1–13, 2018

  22. [30]

    Dif- ferentiable vector graphics rasterization for editing and learning

    Tzu-Mao Li, Michal Luk´ ac, Micha¨ el Gharbi, and Jonathan Ragan-Kelley. Dif- ferentiable vector graphics rasterization for editing and learning. ACM Trans- actions on Graphics (TOG) , 39(6):1–15, 2020. 19

  23. [31]

    Opendr: An approximate differen- tiable renderer

    Matthew M Loper and Michael J Black. Opendr: An approximate differen- tiable renderer. In European Conference on Computer Vision, pages 154–169. Springer, 2014

  24. [32]

    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. ACM siggraph computer graphics , 21(4):163– 169, 1987

  25. [33]

    Modeling, inference and optimization with composable dif- ferentiable procedures

    Dougal Maclaurin. Modeling, inference and optimization with composable dif- ferentiable procedures. PhD thesis, Harvard, 2016

  26. [34]

    Nurbs-based and parametric-based shape optimization with differentiated cad kernel

    Orest Mykhaskiv, Mladen Banovi´ c, Salvatore Auriemma, Pavanakumar Mo- hanamuraly, Andrea Walther, Herve Legrand, and Jens-Dominik M¨ uller. Nurbs-based and parametric-based shape optimization with differentiated cad kernel. Computer-Aided Design and Applications , 15(6):916–926, 2018

  27. [35]

    Level set methods and dynamic implicit surfaces, volume 153

    Stanley Osher and Ronald Fedkiw. Level set methods and dynamic implicit surfaces, volume 153. Springer Science & Business Media, 2006

  28. [36]

    Hyperfun project: A framework for collab- orative multidimensional f-rep modeling

    Alexander Pasko, Valery Adzhiev, Richard Cartwright, Eric Fausett, Anatoly Ossipov, and Vladimir Savchenko. Hyperfun project: A framework for collab- orative multidimensional f-rep modeling. In Eurographics/ACM SIGGRAPH Workshop Implicit Surfaces’ 99 , pages 59–69, 1999

  29. [37]

    Function representation in geometric modeling: concepts, implementation and applications

    Alexander Pasko, Valery Adzhiev, Alexei Sourin, and Vladimir Savchenko. Function representation in geometric modeling: concepts, implementation and applications. The visual computer , 11(8):429–446, 1995

  30. [38]

    Procedural function-based modelling of volumetric mi- crostructures

    Alexander Pasko, Oleg Fryazinov, Turlif Vilbrandt, Pierre-Alain Fayolle, and Valery Adzhiev. Procedural function-based modelling of volumetric mi- crostructures. Graphical Models, 73(5):165–181, 2011

  31. [39]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, and et al. Pytorch: An imperative style, high-performance deep learning library. ArXiv preprint arXiv:1912.01703, 2019

  32. [40]

    Reverse-mode ad in a func- tional framework: Lambda the ultimate backpropagator

    Barak A Pearlmutter and Jeffrey Mark Siskind. Reverse-mode ad in a func- tional framework: Lambda the ultimate backpropagator. ACM Transactions on Programming Languages and Systems (TOPLAS) , 30(2):1–36, 2008

  33. [41]

    Modeling with distance functions

    Inigo Quilez. Modeling with distance functions. http://iquilezles.org/ www/articles/distfunctions/distfunctions.htm

  34. [42]

    Regularized evolution for image classifier architecture search

    Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V Le. Regularized evolution for image classifier architecture search. In Proceedings of the aaai conference on artificial intelligence, volume 33, pages 4780–4789. AAAI, 2019

  35. [43]

    A. Ricci. A Constructive Geometry for Computer Graphics. The Computer Journal, 16(2):157–160, May 1973

  36. [44]

    A Stochastic Approximation Method

    Herbert Robbins and Sutton Monro. A Stochastic Approximation Method. The Annals of Mathematical Statistics , 22(3):400–407, 1951. 20

  37. [45]

    V. L. Rvachev. Methods of Logic Algebra in Mathematical Physics . Naukova Dumka, 1974. In Russian

  38. [46]

    Schmidt, B

    R. Schmidt, B. Wyvill, M. C. Sousa, and J. A. Jorge. Shapeshop: Sketch- based solid modeling with blobtrees. In ACM SIGGRAPH 2007 Courses , SIGGRAPH ’07, page 43, New York, NY, USA, 2007. ACM, Association for Computing Machinery

  39. [47]

    Semi-analytic geometry

    Vadim Shapiro. Semi-analytic geometry. Acta Numerica 2007: Volume 16 , 16:239–303, 2007

  40. [48]

    Match: differentiable material graphs for procedural material capture

    Liang Shi, Beichen Li, Milos Hasan, Kalyan Sunkavalli, Tamy Boubekeur, Radomir Mech, and Wojciech Matusik. Match: differentiable material graphs for procedural material capture. ACM Transactions on Graphics (TOG) , 39(6):1–15, 2020

  41. [49]

    Distance approximations for rasterizing implicit curves

    Gabriel Taubin. Distance approximations for rasterizing implicit curves. ACM Transactions on Graphics, 13:3–42, 1994

  42. [50]

    Tewari, O

    A. Tewari, O. Fried, J. Thies, V. Sitzmann, and et al. State of the art on neural rendering. Computer Graphics Forum, 39(2):701–727, 2020

  43. [51]

    Openfab: A programmable pipeline for multi-material fabrication

    Kiril Vidimce, Szu-Po Wang, Jonathan Ragan-Kelley, and Wojciech Matusik. Openfab: A programmable pipeline for multi-material fabrication. ACM Transactions on Graphics, 32, July 2013

  44. [52]

    Extending the csg tree

    Brian Wyvill, Eric Galin, and Andrew Guy. Extending the csg tree. warp- ing, blending and boolean operations in an implicit surface modeling system. Computer Graphics Forum, 18(2):149–158, 1999. 21

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.