Pith. sign in

REVIEW 3 major objections 3 minor 55 references

ConTiCoM-3D claims single-step point-cloud generation directly in raw point space, using a continuous-time consistency predictor, an analytic velocity target, and Chamfer reconstruction.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

ConTiCoM-3D trains a continuous-time consistency-style model directly on raw 3D point clouds using flow matching plus Chamfer distance, with one- to two-step generation.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection The single-step generator doesn't hold up: the zero-loss premise is unattainable and f_theta(x_T,T) collapses to a posterior mean, but the two-step ODE sampler is a real contribution worth a reviewer's time. the 3 major comments →

arxiv 2509.01492 v1 pith:F6QAJUAQ submitted 2025-09-01 cs.CV

A Continuous-Time Consistency Model for 3D Point Cloud Generation

Sebastian Eilermann , Ren\'e Heesch , Oliver Niggemann This is my paper
classification cs.CV
keywords continuous-time consistency model3D point cloud generationsingle-step samplingflow matchingTrigFlowChamfer distanceShapeNetteacher-free generation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

ConTiCoM-3D sets out to make 3D point-cloud generation as fast as one network evaluation without giving up shape fidelity. It trains a time-conditioned network directly in raw point space, using the TrigFlow forward schedule, an analytic flow-matching velocity target, and a Chamfer-distance reconstruction loss. The central claim is that this combination is the first continuous-time consistency model for point clouds that needs no pre-trained diffusion teacher, no Jacobian-vector products, and no latent autoencoder. If true, it would make real-time generation of high-resolution 3D shapes practical for robotics, AR/VR, and digital content creation, with one- to two-step inference at 0.22–0.42 seconds per shape. On ShapeNet, the paper reports that one- and two-step variants match or beat diffusion and latent-consistency baselines on 1-NNA metrics.

Core claim

The paper claims that point-cloud generation can be reduced to one or two evaluations of a time-conditioned network in raw point space. Using the TrigFlow schedule xt = cos(t)x0 + sin(t)z and the predictor fθ(xt,t) = cos(t)xt − sin(t)σdFθ(xt/σd,t), it proves (Proposition 1) that if Fθ fits the analytic velocity cos(t)z − sin(t)x0, then fθ(xt,t) = x0 for every t, so a single evaluation at t = π/2 maps Gaussian noise to a clean shape. Training combines a flow-matching regression to that analytic velocity with a Chamfer-distance reconstruction term using adaptive time weighting; there is no pre-trained teacher, no Jacobian–vector product, and no latent code. On ShapeNet single-class and 13-clas

What carries the argument

The load-bearing object is the closed-form continuous-time consistency predictor fθ(xt,t) = cos(t)xt − sin(t)σdFθ(xt/σd,t). It couples the TrigFlow forward path xt = cos(t)x0 + sin(t)z with a velocity network Fθ; when Fθ matches the analytic velocity cos(t)z − sin(t)x0, the cos² + sin² identity cancels the noise term exactly and leaves x0. The companion machinery is the training objective: analytic flow-matching regression plus a permutation-invariant Chamfer term, which replaces JVP- or teacher-based consistency supervision.

Load-bearing premise

The whole construction depends on one regression being learnable: from a noisy point cloud, the network must recover which clean shape and which noise produced it, even at the noisiest time where the input is almost pure noise.

What would settle it

Train the model on a dataset of just two distinct shapes. At t = π/2 every training pair (x0, z) has the same network input xt = z, but the regression target cos(t)z − sin(t)x0 = −x0 differs by shape; if the optimizer drives the flow-matching loss down only by predicting the average shape, single-step generation diversity will collapse. Comparing S=1 and S=2 generation diversity on a full class would expose whether the pure-noise endpoint loses information that the two-step retrace restores.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • One network evaluation at t = π/2 can generate a full point cloud; two Euler steps recover most of the additional fidelity, so real-time inference needs only one or two evaluations.
  • Training needs no pre-trained diffusion teacher and no perceptual metric for distillation, removing a major obstacle to consistency models in point space.
  • Because supervision is an analytic velocity regression plus Chamfer distance, no Jacobian–vector products are computed, keeping memory and runtime practical for 2,048-point clouds.
  • Generating in raw point space avoids encoder–decoder compression, so the model does not inherit latent-bottleneck artifacts.
  • On ShapeNet, the one- and two-step variants match or beat the compared diffusion and latent-consistency models on 1-NNA while running at 0.22–0.42 seconds per shape.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because Proposition 1 predicts reconstruction error scaling as sin(t)·εt, one could diagnose the model by plotting |fθ(xt,t) − x0| against t; a sine-shaped envelope would confirm the cancellation mechanism drives fidelity.
  • The method's objective should transfer to conditional or guided generation by conditioning Fθ on a class label, text embedding, or partial scan; the paper mentions this as future work without testing it.
  • The Chamfer term is permutation-invariant and count-agnostic, so the training recipe might extend directly to variable-resolution or partial point clouds without resampling every shape to 2,048 points.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The paper introduces ConTiCoM-3D, a continuous-time consistency model for 3D point cloud generation that operates directly in raw point space. The method uses a TrigFlow forward process x_t = cos(t)x_0 + sin(t)z, a closed-form predictor f_θ(x_t,t)=cos(t)x_t - sin(t)σ_d F_θ(x_t/σ_d,t), and a training objective consisting of an analytic flow-matching loss plus a Chamfer reconstruction loss. The authors claim that this enables teacher-free, JVP-free, latent-free one- to two-step generation, with state-of-the-art 1-NNA results on ShapeNet. The central theoretical statement is Proposition 1, which asserts that if F_θ matches the per-sample analytic velocity, then f_θ(x_t,t)=x_0 for all t.

Significance. If the single-step generation claim were correct, the method would be a meaningful practical contribution: fast, geometry-preserving 3D generation without distillation, Jacobian-vector products, or latent encodings. The paper is clearly written, includes source-code availability, and provides extensive ablations and comparisons. However, the core theoretical mechanism is flawed: the zero-loss premise of Proposition 1 is not satisfiable by any function of the network input, and the training objective optimizes a conditional-mean denoiser rather than a consistency mapping. The reported one-step diversity is therefore not explained by the method as presented. The multi-step Euler sampler in Eq. (16) may still be a plausible flow-matching generator, but that is a different claim and is not the paper's headline contribution.

major comments (3)
  1. [Sec. 4.2, Eq. (12), Proposition 1 and Appendix C] The zero-loss premise of Proposition 1 is unattainable. The regression target in Eq. (12), cos(t)z - sin(t)x0, is not a function of the network input x_t. At t = Tmax = π/2, x_t = z, and z is drawn independently of x0, so the target equals -x0. No function F_θ(x_t/σ_d, t) can equal this per-sample target for all pairs (x0,z). The squared-loss optimum over measurable functions is the conditional expectation, which at t=π/2 is -E[x0]. Consequently Eq. (15) yields f_θ(z,Tmax)=E[x0], the mean shape, not a diverse sample. The proof in Appendix C assumes exactly the impossible equality and therefore does not establish the single-step recovery claim.
  2. [Eq. (10) and the role of the learned velocity] Even if F_θ converges to the best possible approximation, namely the conditional velocity field v(x_t,t)=E[cos(t)z - sin(t)x0 | x_t], the predictor in Eq. (10) reduces to the posterior mean E[x0|x_t], not to the consistency-flow endpoint. Using Tweedie's identity, v = (σ_d²/cos t)∇ log p_t - tan(t)x_t, and substituting into Eq. (10) gives (x_t - sin(t)E[z|x_t])/cos(t) = E[x0|x_t]. This is the MMSE denoiser, which is deterministic and, at t=Tmax, is the constant mean. The Chamfer term in Eq. (13) has the same identifiability problem: it supervises a deterministic map from z to a shape whose target x0 is independent of z. Thus the training objective does not learn a diverse consistency function.
  3. [Sec. 5.2, Tables 1-3 and Fig. 2-3] The empirical one-step results are not supported by the theoretical framework. Given that the objective's expected optimum is the mean shape, the reported low 1-NNA values and the interpolations in Fig. 3 are not predicted by the method as stated. The paper provides no diagnostic to show that the S=1 outputs have nontrivial diversity, that F_θ at Tmax is not approximately constant, or that the network generalizes beyond memorizing finite (z,x0) training pairs. Without such evidence, the central claim of one-step diverse 3D generation is unsubstantiated.
minor comments (3)
  1. [Sec. 3] The 'Adversarial and Autoencoding Approaches' paragraph is duplicated verbatim in Section 3.
  2. [Appendix E, Table 7] The column header 'MMD (EMD)↓' is confusing; MMD is usually reported under CD, and the row values are not clearly labeled.
  3. [Throughout] There are several typographical issues (e.g., 'V AEs', 'F orty-second', inconsistent spacing in citations). These do not affect the technical content but should be cleaned up.

Circularity Check

2 steps flagged

Single-step generation reduces to the training reconstruction by construction: Proposition 1 is an algebraic tautology, and at T=pi/2 the predictor input is pure noise independent of x0, forcing a conditional mean.

specific steps
  1. self definitional [Sec. 4.1 Eqs. 10-12; Appendix C (Proposition 1)]
    "In the ideal zero-loss case, this implies the analytic regression target σdFθ(xt/σd, t) = cos t z− sin t x0. Then, for all t ∈ [0, Tmax], fθ(xt, t) = x0."

    The predictor fθ is parameterized in Eq. 10 as cos(t)xt − sin(t)σd Fθ(xt/σd,t), exactly the form that cancels xt when the FM regression target of Eq. 12 is substituted. Proposition 1 therefore proves an identity: if Fθ equals the analytic velocity, then fθ = x0 by construction. The 'closed-form recovery' is not an independent consequence of the method; it is the zero-loss version of the training objective, so the derivation reduces to the loss definition.

  2. fitted input called prediction [Sec. 4.3 Eq. 15; cf. Eqs. 9, 12, 13]
    "Single-step generation. Sampling requires only one evaluation at T = Tmax: ˆx0 = fθ(xT , T), x T ∼ N(0, σ2 dI)."

    At T = Tmax = π/2, cos T = 0 and sin T = 1, so Eq. 15 becomes fθ(z,T) = −σd Fθ(z/σd,T). The FM target in Eq. 12 at this time is −x0, which does not depend on z, and z is drawn independently of x0. Since the network input z carries no information about x0, the L2-optimal Fθ is the constant −E[x0]; the Chamfer term (Eq. 13) reinforces this deterministic mean reconstruction. Thus the advertised one-step 'generation' is the fitted conditional mean E[x0], not a diverse map from noise to shapes. The prediction is exactly the training reconstruction target by construction.

full rationale

The central derivation is circular in a specific, quotable sense. Equation 10 defines the predictor so that substituting the analytic TrigFlow velocity yields fθ(xt,t)=x0; Proposition 1 is therefore a tautological restatement of the FM loss rather than an independent theoretical guarantee. The single-step sampler (Eq. 15) evaluates this predictor at T=π/2, where xt=z is pure noise independent of the target x0. Because both the FM regression (Eq. 12) and the Chamfer term (Eq. 13) supervise a deterministic function of z to predict x0, the unique attainable solution is the constant mean shape; no finite-capacity network can recover x0 from a variable that does not contain it. Hence the headline 'one-to-two-step generation with diversity' reduces by construction to a fitted reconstruction/mean. The paper's own Appendix C notes error amplification near T but does not address this identifiability failure. The multi-step ODE sampler (Eq. 16) is a legitimate flow-matching generator, and the empirical comparisons against external baselines are not circular, so the paper is not wholly vacuous; however, the single-step claim central to the title and abstract is not supported by the derivation. No load-bearing self-citation was found.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The central claim rests on the assumption that a consistency function mapping pure noise to data exists and is learnable through the proposed losses; this is not established because at t=T the regression target is not identifiable from the input. No new physical or relational entities are posited; the adaptive weighting is a hyperparameter schedule, not an invented entity.

free parameters (2)
  • Adaptive Chamfer weight lambda_CD(t) = [0.1, 0.3]
    Time-dependent weighting between FM and Chamfer losses, chosen by hand (Appendix B.2).
  • Data scale sigma_d = 1.0
    Set to 1.0 following TrigFlow; interacts with point cloud normalization and noise scale (Eq. 9, Appendix B.2).
axioms (4)
  • domain assumption The forward process xt=cos(t)x0+sin(t)z defines a valid probability flow ODE with analytic velocity cos(t)z - sin(t)x0.
    Standard TrigFlow/flow matching assumption used in Eqs. 9-12.
  • ad hoc to paper A neural network F_theta can be trained so that sigma_d F_theta(xt/sigma_d,t) approximates the per-sample velocity cos(t)z - sin(t)x0.
    This is the zero-loss premise of Proposition 1; it is not attainable in general because at t=pi/2, xt=z is independent of x0, so the target is not a function of the input (Appendix C).
  • domain assumption Chamfer distance is an adequate training and evaluation metric for generative point cloud quality.
    Used for LCD (Eq. 13) and for 1-NNA CD evaluation (Section 5.1).
  • domain assumption The learned consistency function is a well-defined deterministic map from noise to data.
    Single-step sampling Eq. 15 assumes f_theta(x_T,T) generates diverse shapes, but the pairing of z and x0 during training is independent, so the map is not determined.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of A Continuous-Time Consistency Model for 3D Point Cloud Generation." pith.science (2026). https://pith.science/paper/F6QAJUAQ

@misc{pith2026250901492,
  author       = {Pith},
  title        = {Pith review of: A Continuous-Time Consistency Model for 3D Point Cloud Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F6QAJUAQ}},
  note         = {Machine review of arXiv:2509.01492}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Fast and accurate 3D shape generation from point clouds is essential for applications in robotics, AR/VR, and digital content creation. We introduce ConTiCoM-3D, a continuous-time consistency model that synthesizes 3D shapes directly in point space, without discretized diffusion steps, pre-trained teacher models, or latent-space encodings. The method integrates a TrigFlow-inspired continuous noise schedule with a Chamfer Distance-based geometric loss, enabling stable training on high-dimensional point sets while avoiding expensive Jacobian-vector products. This design supports efficient one- to two-step inference with high geometric fidelity. In contrast to previous approaches that rely on iterative denoising or latent decoders, ConTiCoM-3D employs a time-conditioned neural network operating entirely in continuous time, thereby achieving fast generation. Experiments on the ShapeNet benchmark show that ConTiCoM-3D matches or outperforms state-of-the-art diffusion and latent consistency models in both quality and efficiency, establishing it as a practical framework for scalable 3D shape generation.

Figures

Figures reproduced from arXiv: 2509.01492 by Oliver Niggemann, Ren\'e Heesch, Sebastian Eilermann.

Figure 1
Figure 1. Figure 1: ConTiCoM-3D models a TrigFlow-inspired forward flow ODE (dashed line) that perturbs a clean point cloud [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Single-class generation results (airplane, chair, car). Top row: samples generated by ConTiCoM-3D with S = 1 step. Bottom row: samples from MLPCM [8]. ConTiCoM-3D produces sharper structures and more consistent global geometry while maintaining diversity, whereas MLPCM exhibits compression artifacts from latent bottlenecks. Quantitative Results [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Latent interpolation results with ConTiCoM-3D. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Further qualitative results with ConTiCoM-3D (S=1) on ShapeNet dataset [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

55 extracted references · 29 canonical work pages · 5 internal anchors

  1. [1]

    Learning representations and generative models for 3d point clouds

    Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning , pages 40–49. PMLR, 2018

    work page 2018

  2. [2]

    How to build a consistency model: Learning flow maps via self-distillation

    Nicholas M Boffi, Michael S Albergo, and Eric Vanden-Eijnden. How to build a consistency model: Learning flow maps via self-distillation. arXiv preprint arXiv:2505.18825, 2025

    arXiv 2025

  3. [3]

    Learning gradient fields for shape generation

    Ruojin Cai, Guandao Yang, Hadar Averbuch-Elor, Zekun Hao, Serge Belongie, Noah Snavely, and Bharath Hariharan. Learning gradient fields for shape generation. In Computer Vision–ECCV 2020 , pages 364–381. Springer, 2020

    work page 2020

  4. [4]

    Shapenet: An information-rich 3d model repository

    Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015

    Pith/arXiv arXiv 2015

  5. [5]

    Sana-sprint: One-step diffusion with continuous-time consistency distillation

    Junsong Chen, Shuchen Xue, Yuyang Zhao, Jincheng Yu, Sayak Paul, Junyu Chen, Han Cai, Song Han, and Enze Xie. Sana-sprint: One-step diffusion with continuous-time consistency distillation. arXiv preprint arXiv:2503.09641, 2025

    arXiv 2025

  6. [6]

    Convergence Of Consistency Model With Multistep Sampling Under General Data Assumptions

    Yiding Chen, Yiyi Zhang, Owen Oertell, and Wen Sun. Convergence of consistency model with multistep sampling under general data assumptions. arXiv preprint arXiv:2505.03194, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  7. [7]

    Improved training technique for latent consistency models

    Quan Dao, Khanh Doan, Di Liu, Trung Le, and Dimitris Metaxas. Improved training technique for latent consistency models. arXiv preprint arXiv:2502.01441, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    Multi-scale latent point consistency models for 3d shape generation

    Bi’an Du, Wei Hu, and Renjie Liao. Multi-scale latent point consistency models for 3d shape generation. In arXiv preprint arXiv:2412.19413, 2024

    arXiv 2024

  9. [9]

    A point set generation network for 3d object reconstruction from a single image

    Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 605–613, 2017

    work page 2017

  10. [10]

    Towards an MLOps Architecture for XAI in Industrial Applications

    Leonhard Faubel, Thomas Woudsma, Leila Methnani, Amir Ghorbani Ghezeljhemeidan, Fabian Buelow, Klaus Schmid, Willem D van Driel, Benjamin Kloepper, Andreas Theodorou, Mohsen Nosratinia, et al. Towards an mlops architecture for xai in industrial applications. arXiv preprint arXiv:2309.12756, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  11. [11]

    Point cloud diffusion models for automatic implant generation

    Paul Friedrich, Julia Wolleb, Florentin Bieder, Florian M Thieringer, and Philippe C Cattin. Point cloud diffusion models for automatic implant generation. In International conference on medical image computing and computer- assisted intervention, pages 112–122. Springer, 2023. 9 ConTiCoM-3D

    work page 2023

  12. [12]

    Multistep consistency models

    Jonathan Heek, Emiel Hoogeboom, and Tim Salimans. Multistep consistency models. arXiv preprint arXiv:2403.06807, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  14. [14]

    Design automation: A conditional vae approach to 3d object generation under conditions

    Michael Hohmann, Sebastian Eilermann, Willi Großmann, and Oliver Niggemann. Design automation: A conditional vae approach to 3d object generation under conditions. In 2024 IEEE 29th International Conference on Emerging Technologies and Factory Automation (ETF A), pages 1–8. IEEE, 2024

    work page 2024

  15. [15]

    Not-so-optimal transport flows for 3d point cloud generation

    Ka-Hei Hui, Chao Liu, Xiaohui Zeng, Chi-Wing Fu, and Arash Vahdat. Not-so-optimal transport flows for 3d point cloud generation. arXiv preprint arXiv:2502.12456, 2025

    Pith/arXiv arXiv 2025

  16. [16]

    Progressive point cloud deconvolution generation network

    Le Hui, Rui Xu, Jin Xie, Jianjun Qian, and Jian Yang. Progressive point cloud deconvolution generation network. In Computer Vision–ECCV 2020, pages 397–413. Springer, 2020

    work page 2020

  17. [17]

    Progressive point cloud deconvolution generation network

    Le Hui, Rui Xu, Jin Xie, Jianjun Qian, and Jian Yang. Progressive point cloud deconvolution generation network. In European conference on computer vision, pages 397–413. Springer, 2020

    work page 2020

  18. [18]

    Consistency Diffusion Models for Single-Image 3D Reconstruction with Priors

    Chenru Jiang, Chengrui Zhang, Xi Yang, Jie Sun, Yifei Zhang, Bin Dong, and Kaizhu Huang. Consistency diffusion models for single-image 3d reconstruction with priors. arXiv preprint arXiv:2501.16737, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  19. [19]

    Softflow: Probabilistic framework for normalizing flow on manifolds

    Hyeongju Kim, Hyeonseung Lee, Woo Hyun Kang, Joun Yeop Lee, and Nam Soo Kim. Softflow: Probabilistic framework for normalizing flow on manifolds. In Advances in Neural Information Processing Systems , volume 33, pages 16388–16397, 2020

    work page 2020

  20. [20]

    Setvae: Learning hierarchical composition for generative modeling of set-structured data

    Jinwoo Kim, Jaehoon Yoo, Juho Lee, and Seunghoon Hong. Setvae: Learning hierarchical composition for generative modeling of set-structured data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15059–15068, 2021

    work page 2021

  21. [21]

    Adam: A method for stochastic optimization

    Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014

    Pith/arXiv arXiv 2014

  22. [22]

    Discrete point flow networks for efficient point cloud generation

    Roman Klokov, Edmond Boyer, and Jakob Verbeek. Discrete point flow networks for efficient point cloud generation. In European Conference on Computer Vision, pages 694–710. Springer, 2020

    work page 2020

  23. [23]

    Sp-gan: Sphere-guided 3d shape generation and manipula- tion

    Ruihui Li, Xianzhi Li, Ka-Hei Hui, and Chi-Wing Fu. Sp-gan: Sphere-guided 3d shape generation and manipula- tion. ACM Transactions on Graphics (TOG), 40(4):1–12, 2021

    work page 2021

  24. [24]

    Generalized deep 3d shape prior via part-discretized diffusion process

    Yuhan Li, Yishun Dou, Xuanhong Chen, Bingbing Ni, Yilin Sun, Yutian Liu, and Fuzhen Wang. Generalized deep 3d shape prior via part-discretized diffusion process. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16784–16794, 2023

    work page 2023

  25. [25]

    Flow matching for generative modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022

    Pith/arXiv arXiv 2022

  26. [26]

    Treegan: Syntax-aware sequence generation with generative adversarial networks

    Xinyue Liu, Xiangnan Kong, Lei Liu, and Kuorong Chiang. Treegan: Syntax-aware sequence generation with generative adversarial networks. In 2018 IEEE International Conference on Data Mining (ICDM) , pages 1140–1145. IEEE, 2018

    work page 2018

  27. [27]

    Meshdiffusion: Score-based generative 3d mesh modeling

    Zhen Liu, Yao Feng, Michael J Black, Derek Nowrouzezahrai, Liam Paull, and Weiyang Liu. Meshdiffusion: Score-based generative 3d mesh modeling. arXiv preprint arXiv:2303.08133, 2023

    Pith/arXiv arXiv 2023

  28. [28]

    Point-voxel cnn for efficient 3d deep learning

    Zhijian Liu, Haotian Tang, Yujun Lin, and Song Han. Point-voxel cnn for efficient 3d deep learning. In Advances in Neural Information Processing Systems , 2019

    work page 2019

  29. [29]

    Simplifying, stabilizing and scaling continuous-time consistency models

    Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models. arXiv preprint arXiv:2410.11081, 2024

    Pith/arXiv arXiv 2024

  30. [30]

    Manicm: Real-time 3d diffusion policy via consistency model for robotic manipulation

    Guanxing Lu, Zifeng Gao, Tianxing Chen, Wenxun Dai, Ziwei Wang, Wenbo Ding, and Yansong Tang. Manicm: Real-time 3d diffusion policy via consistency model for robotic manipulation. arXiv preprint arXiv:2406.01586, 2024

    Pith/arXiv arXiv 2024

  31. [31]

    Diffusion probabilistic models for 3d point cloud generation

    Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 2837–2845, 2021

    work page 2021

  32. [32]

    Dit-3d: Exploring plain diffusion transformers for 3d shape generation

    Shentong Mo, Enze Xie, Ruihang Chu, Lanqing Hong, Matthias Niessner, and Zhenguo Li. Dit-3d: Exploring plain diffusion transformers for 3d shape generation. In Advances in Neural Information Processing Systems , volume 36, pages 67960–67971, 2023

    work page 2023

  33. [33]

    Dit-3d: Exploring plain diffusion transformers for 3d shape generation

    Shentong Mo, Enze Xie, Ruihang Chu, Lanqing Hong, Matthias Niessner, and Zhenguo Li. Dit-3d: Exploring plain diffusion transformers for 3d shape generation. Advances in neural information processing systems , 36:67960–67971, 2023. 10 ConTiCoM-3D

    work page 2023

  34. [34]

    Shape as points: A differentiable poisson solver

    Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. Advances in Neural Information Processing Systems , 34:13032–13044, 2021

    work page 2021

  35. [35]

    Convolutional occupancy networks

    Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 523–540. Springer, 2020

    2020

  36. [36]

    A Generative Neural Network Approach for 3D Multi-Criteria Design Generation and Optimization of an Engine Mount for an Unmanned Air Vehicle

    Christoph Petroll, Sebastian Eilermann, Philipp Hoefer, and Oliver Niggemann. A generative neural network approach for 3d multi-criteria design generation and optimization of an engine mount for an unmanned air vehicle. arXiv preprint arXiv:2311.03414, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  37. [37]

    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, pages 652–660, 2017

    2017

  38. [38]

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space

    Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems , 30, 2017

    2017

  39. [39]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning , pages 8748–8763. PmLR, 2021

    2021

  40. [40]

    The earth mover’s distance as a metric for image retrieval

    Yossi Rubner, Carlo Tomasi, and Leonidas J Guibas. The earth mover’s distance as a metric for image retrieval. International journal of computer vision , 40(2):99–121, 2000

    work page 2000

  41. [41]

    Align your flow: Scaling continuous-time flow map distillation

    Amirmojtaba Sabour, Sanja Fidler, and Karsten Kreis. Align your flow: Scaling continuous-time flow map distillation. arXiv preprint arXiv:2506.14603, 2025

    Pith/arXiv arXiv 2025

  42. [42]

    Progressive distillation for fast sampling of diffusion models

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022

    Pith/arXiv arXiv 2022

  43. [43]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020

    Pith/arXiv arXiv 2010

  44. [44]

    Improved techniques for training consistency models

    Yang Song and Prafulla Dhariwal. Improved techniques for training consistency models. arXiv preprint arXiv:2310.14189, 2023

    Pith/arXiv arXiv 2023

  45. [45]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023

    Pith/arXiv arXiv 2023

  46. [46]

    Generative modeling by estimating gradients of the data distribution

    Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019

    2019

  47. [47]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020

    Pith/arXiv arXiv 2011

  48. [48]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems , 30, 2017

    2017

  49. [49]

    Fast point cloud generation with straight flows

    Lemeng Wu, Dilin Wang, Chengyue Gong, Xingchao Liu, Yunyang Xiong, Rakesh Ranjan, Raghuraman Krishnamoorthi, Vikas Chandra, and Qiang Liu. Fast point cloud generation with straight flows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 9445–9454, 2023

    work page 2023

  50. [50]

    Pointflow: 3d point cloud generation with continuous normalizing flows

    Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4541–4550, 2019

    work page 2019

  51. [51]

    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 , volume 35, pages 10021–10039, 2022

    work page 2022

  52. [52]

    Learning to Generate 3D Shapes with Generative Cellular Automata

    Dongsu Zhang, Changwoon Choi, Jeonghwan Kim, and Young Min Kim. Learning to generate 3d shapes with generative cellular automata. arXiv preprint arXiv:2103.04130, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  53. [53]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586–595, 2018

    2018

  54. [54]

    Inverse flow and consistency models

    Yuchen Zhang and Jian Zhou. Inverse flow and consistency models. In F orty-second International Conference on Machine Learning, 2025

    work page 2025

  55. [55]

    3d shape generation and completion through point-voxel diffusion

    Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 5826–5835, 2021. 11 ConTiCoM-3D A Inferencing As shown in Table 4, our ConTiCoM-3D does not achieve the lowest runtimes reported by purely distilled one-step models such a...

    work page 2021

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.