pith. machine review for the scientific record. sign in

arxiv: 2604.24586 · v1 · submitted 2026-04-27 · 💻 cs.CV

Recognition: unknown

Point-MF: One-step Point Cloud Generation from a Single Image via Mean Flows

Keiji Yanai, Yuta Baba

Authors on Pith no claims yet

Pith reviewed 2026-05-08 04:24 UTC · model grok-4.3

classification 💻 cs.CV
keywords single-image point cloud reconstructionmean flowone-step generationdiffusion transformer3D reconstructionShapeNet-R2N2Pix3Dlow-NFE inference
0
0 comments X

The pith

A mean-flow model reconstructs complete 3D point clouds from one RGB image using a single network evaluation.

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

The paper proposes a framework that learns the mean velocity field directly in point-cloud space rather than through iterative denoising or latent encodings. By pairing an interval-conditioned Diffusion Transformer with an auxiliary anchor loss on the predicted clean points, the method performs reconstruction in one forward pass. This removes the need for dozens of denoising steps while preserving the ability to infer occluded geometry. A sympathetic reader would care because single-image 3D reconstruction becomes fast enough for interactive or real-time use without sacrificing output fidelity on standard benchmarks.

Core claim

Point-MF operates directly in point-cloud space to learn the mean velocity field and enables one-step reconstruction with a single network function evaluation without relying on VAE-based latent representations. It employs a Diffusion Transformer conditioned on frozen image features via a token adapter together with explicit interval conditioning, and introduces the Denoised Space Anchor loss on the induced denoised-space estimate to stabilize large interval jumps and suppress outliers and density artifacts. On ShapeNet-R2N2 and Pix3D the resulting model produces high-quality point clouds at millisecond latency while matching the quality-speed trade-off of multi-step diffusion baselines.

What carries the argument

The mean velocity field predicted by the interval-conditioned Diffusion Transformer, regularized by the Denoised Space Anchor set-distance loss on the denoised-space estimate x_theta.

If this is right

  • Reconstruction runs at millisecond latency instead of requiring many denoising iterations.
  • Complete point clouds are generated directly in ambient space without a separate VAE encoder-decoder stage.
  • Quality remains competitive with slower iterative methods on both synthetic and real-image benchmarks.
  • The same architecture and loss combination supports large-step sampling while keeping point distributions clean.

Where Pith is reading between the lines

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

  • The one-step formulation may transfer to other dense 3D outputs such as meshes or implicit surfaces once appropriate distance losses are defined.
  • Because inference cost is now constant and low, the approach could be embedded in mobile or edge pipelines that currently avoid diffusion-style 3D models.
  • If the mean-flow velocity field proves stable under further architectural scaling, training objectives for 3D generative models could be simplified by dropping multi-step sampling schedules.

Load-bearing premise

The Diffusion Transformer with interval conditioning plus the Denoised Space Anchor loss can stabilize large interval jumps in raw point-cloud space without creating density artifacts or outliers that degrade final quality.

What would settle it

If the one-step Point-MF outputs on ShapeNet-R2N2 show measurably higher Chamfer distance or visibly more outliers and density irregularities than the multi-step diffusion baselines, the central claim would be falsified.

Figures

Figures reproduced from arXiv: 2604.24586 by Keiji Yanai, Yuta Baba.

Figure 1
Figure 1. Figure 1: Comparison of single-image point cloud reconstruction methods. Representative diffusion-model-based methods (a) often rely on camera poses during both training and inference, as well as many iterative sampling steps. In contrast, Point-MF (b) does not require camera poses during training or inference, directly predicts the mean velocity field without a VAE, and enables single-step reconstruction with 1-NFE… view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of Mean Flow in image generation. Using the mean velocity field u(xt, r, t), ob￾tained by integrating the Flow Matching velocity field v(xt, t), en￾ables a large jump over the interval [t, r]. Here, t is closer to the noise side and r is closer to the data side. Mean Flow. In contrast to Flow Match￾ing, which learns the instantaneous velocity field v(xt, t), Mean Flow (MF) [11] directly learns… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of Point-MF. Given a single input image, a frozen DINOv3 encoder extracts a global feature and a sequence of patch features. The global feature is com￾bined with the time embedding t and interval-length embedding dt = t − r, and is used to condition the DiT through AdaLN-Zero. The patch features are refined by a lightweight Post-MHSA Adapter and then used as the context for Cross-Attention. The Di… view at source ↗
Figure 4
Figure 4. Figure 4: Structure of a DiT block. Point tokens are updated sequen￾tially by Self-Attention, Cross￾Attention, and an FFN. In each branch, RMS-normalized features are modulated in an AdaLN-Zero manner using scale, shift, and gate parameters generated from the conditioning vector obtained by summing et(t), edt(dt), and eimg. The image context Zctx is fed into Cross-Attention. Meanwhile, Zctx is passed through an adap… view at source ↗
Figure 5
Figure 5. Figure 5: Conceptual illustration of Denoised Space Anchor (DSA). Mean Flow performs a large temporal update from the state xt at time t using the mean velocity field uθ(xt, r, t | c). However, the reconstructed point cloud xθ(xt, r, t | c) obtained by ex￾trapolating this prediction to time 0 may deviate from the data space. To address this issue, we impose an auxiliary set-distance-based loss between the reconstruc… view at source ↗
Figure 6
Figure 6. Figure 6 view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative results on Pix3D. ble 1, where Point-MF achieves consistently strong EMD scores. In contrast, RGB2point, which directly minimizes Chamfer Distance, often yields sharp out￾lines due to strong attraction of points toward the surface, but can exhibit local density imbalance or concentration of points around particular parts. On Pix3D, Point-MF is able to reconstruct challenging structures such as … view at source ↗
read the original abstract

Single-image point cloud reconstruction must infer complete 3D geometry, including occluded parts, from a single RGB image. While diffusion-based reconstructors achieve high accuracy, they typically require many denoising iterations, resulting in slow and expensive inference. We propose Point-MF, a Mean-Flow-based framework for low-NFE single-image point cloud reconstruction that couples a Mean-Flow-compatible architecture with an auxiliary loss. Specifically, Point-MF operates directly in point-cloud space to learn the mean velocity field and enables one-step reconstruction with a single network function evaluation (1-NFE), without relying on VAE-based latent representations. To make Mean Flow effective under large interval jumps, Point-MF employs a Diffusion Transformer tailored to the Mean-Flow setting, conditioned on frozen DINOv3 image features via a lightweight token adapter and equipped with explicit interval/time conditioning. Moreover, we introduce Denoised Space Anchor, a set-distance auxiliary loss on the denoised-space estimate $x_\theta$ induced by the predicted velocity field, to stabilize large-step generation and reduce outliers and density artifacts. On ShapeNet-R2N2 and Pix3D, Point-MF strikes a strong balance between reconstruction quality and inference speed compared to multi-step diffusion baselines and competitive feedforward models, while generating high-quality point clouds with millisecond-level latency.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The paper proposes Point-MF, a Mean-Flow framework for single-image point cloud reconstruction that learns the mean velocity field directly in point-cloud space. It uses a Diffusion Transformer conditioned on frozen DINOv3 image features via a token adapter, with explicit interval conditioning, plus a new Denoised Space Anchor set-distance auxiliary loss on the induced x_θ to stabilize 1-NFE generation. The method claims to achieve high-quality reconstruction on ShapeNet-R2N2 and Pix3D with only a single network evaluation, offering a favorable quality-speed trade-off versus multi-step diffusion baselines and competitive feedforward models, without VAE latents.

Significance. If the central claims are substantiated, the work would represent a meaningful advance in efficient 3D reconstruction by reducing inference to a single forward pass while preserving reconstruction fidelity. The combination of mean-flow velocity prediction with point-cloud-specific architecture and auxiliary regularization is a targeted extension that could enable real-time applications; the avoidance of latent-space representations is a clear design choice that simplifies the pipeline.

major comments (3)
  1. [Method (Denoised Space Anchor) and Experiments] The Denoised Space Anchor auxiliary loss (described in the method section as a set-distance term on x_θ) is presented as essential for preventing outliers and density artifacts under large interval jumps. However, no ablation removing this loss is reported, so it remains unclear whether the mean-velocity predictor with interval conditioning alone can produce clean point clouds at interval=1 or whether the auxiliary term is load-bearing for the 1-NFE claim. Because the loss is a global unordered-set metric, it may be satisfied while local density inconsistencies persist; an ablation table comparing Chamfer/EMD with and without the loss at 1-NFE would directly test this.
  2. [Abstract and Experiments] The abstract and method claim competitive reconstruction quality and millisecond latency on ShapeNet-R2N2 and Pix3D, yet the provided summary supplies no numerical values, baseline tables, or error analysis. Without explicit metrics (e.g., CD, EMD, F-score) and direct comparisons to multi-step diffusion and feedforward models, the asserted quality-speed trade-off cannot be verified as load-bearing for the headline result.
  3. [Architecture] The Diffusion Transformer with interval conditioning is asserted to make Mean Flow effective for large jumps without VAE latents. More detail is needed on the precise encoding of the interval scalar, its injection into the DiT blocks, and any architectural modifications required to keep the velocity field stable at interval=1; without this, reproducibility of the 1-NFE capability is limited.
minor comments (1)
  1. [Abstract] The abstract repeatedly uses qualitative phrases such as 'strong balance' and 'high-quality' without accompanying numbers; adding even a single-line summary of key metrics would improve immediate readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below, providing clarifications, additional details, and indicating where the manuscript will be revised to incorporate the suggestions.

read point-by-point responses

  1. Referee: [Method (Denoised Space Anchor) and Experiments] The Denoised Space Anchor auxiliary loss (described in the method section as a set-distance term on x_θ) is presented as essential for preventing outliers and density artifacts under large interval jumps. However, no ablation removing this loss is reported, so it remains unclear whether the mean-velocity predictor with interval conditioning alone can produce clean point clouds at interval=1 or whether the auxiliary term is load-bearing for the 1-NFE claim. Because the loss is a global unordered-set metric, it may be satisfied while local density inconsistencies persist; an ablation table comparing Chamfer/EMD with and without the loss at 1-NFE would directly test this.

    Authors: We agree that an explicit ablation is required to establish whether the auxiliary loss is load-bearing for stable 1-NFE generation. In the revised manuscript we have added a dedicated ablation study (new Table 4 in Section 4.3) that reports Chamfer Distance and EMD on ShapeNet-R2N2 for the complete model versus the identical architecture trained without the Denoised Space Anchor loss, all evaluated at interval=1. The table also includes nearest-neighbor density statistics to address the concern that a global set-distance metric might mask local inconsistencies. We will further include qualitative point-cloud visualizations in the supplement showing the increase in outliers when the loss is removed. These additions directly test and substantiate the necessity of the term for the one-step claim. revision: yes

  2. Referee: [Abstract and Experiments] The abstract and method claim competitive reconstruction quality and millisecond latency on ShapeNet-R2N2 and Pix3D, yet the provided summary supplies no numerical values, baseline tables, or error analysis. Without explicit metrics (e.g., CD, EMD, F-score) and direct comparisons to multi-step diffusion and feedforward models, the asserted quality-speed trade-off cannot be verified as load-bearing for the headline result.

    Authors: The experiments section of the original manuscript already contains the requested quantitative material: Tables 1 and 2 report CD, EMD, and F-score on both ShapeNet-R2N2 and Pix3D together with direct comparisons against multi-step diffusion baselines (Point-E, Diff3D) and feedforward models (Pixel2Point, OccNet), plus measured inference latency in milliseconds. To make these results immediately visible from the abstract, we have revised the abstract to include representative numerical values and a concise statement of the quality-speed trade-off. We have also added a short error-analysis paragraph in Section 4.2 that reports standard deviations across the test sets. revision: yes

  3. Referee: [Architecture] The Diffusion Transformer with interval conditioning is asserted to make Mean Flow effective for large jumps without VAE latents. More detail is needed on the precise encoding of the interval scalar, its injection into the DiT blocks, and any architectural modifications required to keep the velocity field stable at interval=1; without this, reproducibility of the 1-NFE capability is limited.

    Authors: We have expanded Section 3.2 with a new subsection (3.2.1) that supplies the requested implementation details. The normalized interval scalar is encoded with a 128-dimensional sinusoidal embedding and concatenated to the time embedding; the combined vector is injected via adaLN-Zero modulation applied to the query/key projections and the MLP layers inside every DiT block. To ensure stability at interval=1 we added a residual bypass around the velocity head and applied a small amount of Gaussian noise to the interval embedding during training. These choices, together with the exact layer dimensions and pseudocode, are now documented in the main text and an appendix figure, enabling full reproducibility of the 1-NFE results. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation; empirical method with independent components

full rationale

The paper introduces a Mean-Flow framework with a tailored Diffusion Transformer, interval conditioning, and a new Denoised Space Anchor auxiliary loss to enable 1-NFE point cloud reconstruction. These elements are presented as novel architectural and loss choices motivated by the need to stabilize large interval jumps, not as redefinitions or renamings of prior results. Claims of quality-speed balance are supported by empirical evaluation on ShapeNet-R2N2 and Pix3D rather than any mathematical derivation that reduces outputs to inputs by construction. No self-citation chains, fitted parameters renamed as predictions, or ansatzes smuggled via prior work appear in the provided text.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Ledger is necessarily incomplete because only the abstract is available; no explicit free parameters, background axioms, or invented entities beyond the named loss are detailed.

axioms (1)
  • domain assumption Mean velocity field can be learned directly in point-cloud space for large interval jumps when conditioned on image features
    Central premise of operating the model in raw point space rather than latent space.
invented entities (1)
  • Denoised Space Anchor no independent evidence
    purpose: Auxiliary set-distance loss on the denoised-space estimate to stabilize one-step generation
    Newly introduced loss term whose independent evidence is not provided in the abstract.

pith-pipeline@v0.9.0 · 5534 in / 1372 out tokens · 59871 ms · 2026-05-08T04:24:17.690466+00:00 · methodology

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

74 extracted references · 8 canonical work pages · 5 internal anchors

  1. [1]

    In: Proc

    An, J., Yang, Z., Wang, J., Li, L., Liu, Z., Wang, L., Luo, J.: Bring metric functions into diffusion models. In: Proc. of the International Joint Conference on Artificial Intelligence. pp. 578–586 (2024) 4

    2024

  2. [2]

    In: Proc

    Boss, M., Huang, Z., Vasishta, A., Jampani, V.: Sf3d: Stable fast 3d mesh recon- struction with uv-unwrapping and illumination disentanglement. In: Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16240– 16250 (2025) 2

    2025

  3. [3]

    of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Boulch,A.,Marlet,R.:Poco:Pointconvolutionforsurfacereconstruction.In:Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6302–6314 (2022) 1

    2022

  4. [4]

    In: Proc

    Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proc. of the IEEE/CVF International Conference on Computer Vision. pp. 9650–9660 (2021) 6

    2021

  5. [5]

    ShapeNet: An Information-Rich 3D Model Repository

    Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., Yu, F.: ShapeNet: An Information-Rich 3D Model Repository. arXiv:1512.03012 (2015) 10

    work page internal anchor Pith review arXiv 2015

  6. [6]

    In: Proc

    Chen, H., Gu, J., Chen, A., Tian, W., Tu, Z., Liu, L., Su, H.: Single-stage diffusion nerf: A unified approach to 3d generation and reconstruction. In: Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2416– 2425 (2023) 2

    2023

  7. [7]

    In: Proc

    Chou, G., Bahat, Y., Heide, F.: Diffusion-sdf: Conditional generative modeling of signed distance functions. In: Proc. of the IEEE/CVF International Conference on Computer Vision. pp. 2262–2272 (2023) 2

    2023

  8. [8]

    In: Proc

    Choy, C.B., Xu, D., Gwak, J., Chen, K., Savarese, S.: 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In: Proc. of the European Con- ference on Computer Vision. pp. 628–644 (2016) 2, 3, 10

    2016

  9. [9]

    In: arXiv preprint arXiv:2412.19413 (2024) 2

    Du, B., Hu, W., Liao, R.: Multi-scale latent point consistency models for 3d shape generation. In: arXiv preprint arXiv:2412.19413 (2024) 2

    work page arXiv 2024

  10. [10]

    In: Proc

    Frans,K.,Hafner,D.,Levine,S.,Abbeel,P.:Onestepdiffusionviashortcutmodels. In: Proc. of the International Conference on Learning Representations (2025) 4

    2025

  11. [11]

    In: Advances in Neural Information Processing Systems

    Geng,Z.,Deng,M.,Bai,X.,Kolter,J.Z.,He,K.:Meanflowsforone-stepgenerative modeling. In: Advances in Neural Information Processing Systems. vol. 38 (2025) 2, 4, 5, 20, 21

    2025

  12. [12]

    Improved Mean Flows: On the Challenges of Fastforward Generative Models

    Geng, Z., Lu, Y., Wu, Z., Shechtman, E., Kolter, J.Z., He, K.: Improved mean flows: On the challenges of fastforward generative models. arXiv preprint arXiv:2512.02012 (2025) 4, 14

    work page internal anchor Pith review arXiv 2025

  13. [13]

    In: Proc

    Geng, Z., Pokle, A., Luo, W., Lin, J., Kolter, J.Z.: Consistency models made easy. In: Proc. of the International Conference on Learning Representations (2025) 4

    2025

  14. [14]

    In: Proc

    Graham, B.: Sparse 3d convolutional neural networks. In: Proc. of the British Machine Vision Conference. pp. 150–1 (2015) 2

    2015

  15. [15]

    Advances in Neural Information Processing Systems33, 6840–6851 (2020) 4

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems33, 6840–6851 (2020) 4

    2020

  16. [16]

    In: NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications (2021) 20

    Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications (2021) 20

    2021

  17. [17]

    In: Proc

    Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: Lrm: Large reconstruction model for single image to 3d. In: Proc. of the International Conference on Learning Representations (2024) 2 16 Y. Baba and K. Yanai

    2024

  18. [18]

    In: Proc

    Huang, Z., Boss, M., Vasishta, A., Rehg, J.M., Jampani, V.: Spar3d: Stable point- aware reconstruction of 3d objects from single images. In: Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16860–16870 (2025) 1

    2025

  19. [19]

    In: Proc

    Huang, Z., Jampani, V., Thai, A., Li, Y., Stojanov, S., Rehg, J.M.: Shapeclipper: Scalable 3d shape learning from single-view images via geometric and clip-based consistency. In: Proc. of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition. pp. 12912–12922 (2023) 2

    2023

  20. [20]

    In: Proc

    Huang, Z., Johnson, J., Debnath, S., Rehg, J.M., Wu, C.Y.: Pointinfinity: Resolution-invariant point diffusion models. In: Proc. of the IEEE/CVF Inter- national Conference on Computer Vision. pp. 10050–10060 (2024) 2

    2024

  21. [21]

    Insafutdinov, E., Dosovitskiy, A.: Unsupervised learning of shape and pose with differentiablepointclouds.In:AdvancesinNeuralInformationProcessingSystems. vol. 31 (2018) 3, 11

    2018

  22. [22]

    In: Proc

    Jignasu, A., Balu, A., Sarkar, S., Hegde, C., Ganapathysubramanian, B., Krish- namurthy, A.: Sdfconnect: Neural implicit surface reconstruction of a sparse point cloud with topological constraints. In: Proc. of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition Workshops. pp. 5271–5279 (2024) 1

    2024

  23. [23]

    Shap-e: Generat- ing conditional 3d implicit functions

    Jun, H., Nichol, A.: Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463 (2023) 2

    work page arXiv 2023

  24. [24]

    In: Advances in Neural Information Processing Systems

    Karras,T.,Aittala,M.,Aila,T.,Laine,S.:Elucidatingthedesignspaceofdiffusion- based generative models. In: Advances in Neural Information Processing Systems. vol. 35, pp. 26565–26577 (2022) 4

    2022

  25. [25]

    ACM Transactions on Graphics42(4), 139–1 (2023) 2, 3

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G., et al.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics42(4), 139–1 (2023) 2, 3

    2023

  26. [26]

    In: Proc

    Kim, D., Lai, C.H., Liao, W.H., Murata, N., Takida, Y., Uesaka, T., He, Y., Mit- sufuji, Y., Ermon, S.: Consistency trajectory models: Learning probability flow ode trajectory of diffusion. In: Proc. of the International Conference on Learning Representations (2024) 4

    2024

  27. [27]

    In: Proc

    L Navaneet, K., Mandikal, P., Jampani, V., Babu, V.: Differ: Moving beyond 3d reconstruction with differentiable feature rendering. In: Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (2019) 11

    2019

  28. [28]

    In: Proc

    Lan, Y., Hong, F., Yang, S., Zhou, S., Meng, X., Dai, B., Pan, X., Loy, C.C.: ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In: Proc. of the European Conference on Computer Vision. pp. 112–130. Springer (2024) 2

    2024

  29. [29]

    In: Proc

    Lan, Y., Zhou, S., Lyu, Z., Hong, F., Yang, S., Dai, B., Pan, X., Loy, C.C.: Gaus- sianAnything: Interactive point cloud latent diffusion for 3D generation. In: Proc. of the International Conference on Learning Representations (2025) 1, 7

    2025

  30. [30]

    In: Proc

    Lee, J.J., Benes, B.: Rgb2point: 3d point cloud generation from single rgb im- ages. In: Proc. of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2952–2962 (2025) 2, 3, 11, 12, 23

    2025

  31. [31]

    In: Proc

    Lee, K., Yu, S., Shin, J.: Decoupled meanflow: Turning flow models into flow maps for accelerated sampling. In: Proc. of the International Conference on Learning Representations (2026) 4

    2026

  32. [32]

    Advances in Neural Information Processing Systems37, 63082–63109 (2024) 4

    Lee, S., Lin, Z., Fanti, G.: Improving the training of rectified flows. Advances in Neural Information Processing Systems37, 63082–63109 (2024) 4

    2024

  33. [33]

    In: Proc

    Lipman, Y., Chen, R.T., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. In: Proc. of the International Conference on Learning Repre- sentations (2023) 2, 4, 5 Point-MF 17

    2023

  34. [34]

    In: Proc

    Liu, M., Shi, R., Chen, L., Zhang, Z., Xu, C., Wei, X., Chen, H., Zeng, C., Gu, J., Su, H.: One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. In: Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10072–10083 (2024) 2

    2024

  35. [35]

    In: Proceedings of the IEEE/CVF inter- national conference on computer vision

    Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero- 1-to-3: Zero-shot one image to 3d object. In: Proceedings of the IEEE/CVF inter- national conference on computer vision. pp. 9298–9309 (2023) 2

    2023

  36. [36]

    In: Proc

    Liu, X., Gong, C., Liu, Q.: Flow straight and fast: Learning to generate and transfer data with rectified flow. In: Proc. of the International Conference on Learning Representations (2023) 4

    2023

  37. [37]

    In: Proc

    Liu, Y., Lin, C., Zeng, Z., Long, X., Liu, L., Komura, T., Wang, W.: Syncdreamer: Generating multiview-consistent images from a single-view image. In: Proc. of the International Conference on Learning Representations (2024) 2

    2024

  38. [38]

    In: Proc

    Lu, C., Song, Y.: Simplifying, stabilizing and scaling continuous-time consistency models. In: Proc. of the International Conference on Learning Representations (2025) 4

    2025

  39. [39]

    One-step Latent-free Image Generation with Pixel Mean Flows

    Lu, Y., Lu, S., Sun, Q., Zhao, H., Jiang, Z., Wang, X., Li, T., Geng, Z., He, K.: One-step latent-free image generation with pixel mean flows. In: arXiv preprint arXiv:2601.22158 (2026) 4, 14

    work page internal anchor Pith review arXiv 2026

  40. [40]

    In: Proc

    Luo, S., Hu, W.: Diffusion probabilistic models for 3d point cloud generation. In: Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2837–2845 (2021) 3

    2021

  41. [41]

    In: Proc

    Lyu, Z., Wang, J., An, Y., Zhang, Y., Lin, D., Dai, B.: Controllable mesh gen- eration through sparse latent point diffusion models. In: Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 271–280 (2023) 2

    2023

  42. [42]

    In: Proc

    Ma, B., Han, Z., Liu, Y.S., Zwicker, M.: Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. In: Proc. of the International Conference on Machine Learning. vol. 139 (2021) 2

    2021

  43. [43]

    In: Proc

    Melas-Kyriazi, L., Rupprecht, C., Vedaldi, A.: Pc2: Projection-conditioned point cloud diffusion for single-image 3d reconstruction. In: Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12923–12932 (2023) 2, 3, 11, 12, 23

    2023

  44. [44]

    In: Proc

    Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4460–4470 (2019) 3

    2019

  45. [45]

    Advances in Neural Information Processing Systems36, 67960–67971 (2023) 3

    Mo, S., Xie, E., Chu, R., Hong, L., Niessner, M., Li, Z.: Dit-3d: Exploring plain diffusion transformers for 3d shape generation. Advances in Neural Information Processing Systems36, 67960–67971 (2023) 3

    2023

  46. [46]

    In: Proc

    Mu, Y., Zuo, X., Guo, C., Wang, Y., Lu, J., Wu, X., Xu, S., Dai, P., Yan, Y., Cheng, L.: Gsd: View-guided gaussian splatting diffusion for 3d reconstruction. In: Proc. of the European Conference on Computer Vision. pp. 55–72 (2024) 3

    2024

  47. [47]

    In: Proc

    Murez, Z., Van As, T., Bartolozzi, J., Sinha, A., Badrinarayanan, V., Rabinovich, A.: Atlas: End-to-end 3d scene reconstruction from posed images. In: Proc. of the European Conference on Computer Vision. pp. 414–431 (2020) 2, 3

    2020

  48. [48]

    In: Proc

    Navaneet, K., Mathew, A., Kashyap, S., Hung, W.C., Jampani, V., Babu, R.V.: From image collections to point clouds with self-supervised shape and pose net- works. In: Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1132–1140 (2020) 10, 11

    2020

  49. [49]

    Point-E: A System for Generating 3D Point Clouds from Complex Prompts

    Nichol, A., Jun, H., Dhariwal, P., Mishkin, P., Chen, M.: Point-e: A system for generating 3d point clouds from complex prompts. arXiv:2212.08751 (2022) 2, 3, 8 18 Y. Baba and K. Yanai

    work page internal anchor Pith review arXiv 2022

  50. [50]

    In: Proc

    Park,J.J.,Florence,P.,Straub,J.,Newcombe,R.,Lovegrove,S.:Deepsdf:Learning continuous signed distance functions for shape representation. In: Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 165– 174 (2019) 3

    2019

  51. [51]

    In: Proc

    Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proc. of the IEEE/CVF International Conference on Computer Vision. pp. 4195–4205 (2023) 8

    2023

  52. [52]

    Advances in Neural Information Processing Systems30(2017) 2

    Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learn- ing on point sets in a metric space. Advances in Neural Information Processing Systems30(2017) 2

    2017

  53. [53]

    In: Proc

    Rong, Y., Zhou, H., Xia, K., Mei, C., Wang, J., Lu, T.: Repkpu: Point cloud upsampling with kernel point representation and deformation. In: Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21050– 21060 (2024) 11, 14

    2024

  54. [54]

    In: Advances in Neural Information Processing Systems

    Sharifipour, S., Casado, C.Á., Sabokrou, M., López, M.B.: APML: adaptive prob- abilistic matching loss for robust 3d point cloud reconstruction. In: Advances in Neural Information Processing Systems. vol. 38 (2025) 9

    2025

  55. [55]

    DINOv3

    Siméoni, O., Vo, H.V., Seitzer, M., Baldassarre, F., Oquab, M., Jose, C., Khalidov, V., Szafraniec, M., Yi, S., Ramamonjisoa, M., et al.: Dinov3. arXiv:2508.10104 (2025) 6

    work page internal anchor Pith review arXiv 2025

  56. [56]

    In: Proc

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsuper- vised learning using nonequilibrium thermodynamics. In: Proc. of the International Conference on Machine Learning. pp. 2256–2265 (2015) 4

    2015

  57. [57]

    In: Proc

    Song, Y., Dhariwal, P.: Improved techniques for training consistency models. In: Proc. of the International Conference on Learning Representations (2024) 4

    2024

  58. [58]

    In: Proc

    Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. In: Proc. of the International Conference on Machine Learning (2023) 4

    2023

  59. [59]

    Advances in Neural Information Processing Systems32(2019) 4

    Song, Y., Ermon, S.: Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems32(2019) 4

    2019

  60. [60]

    In: Proc

    Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score- based generative modeling through stochastic differential equations. In: Proc. of the International Conference on Learning Representations (2021) 4

    2021

  61. [61]

    In: Proc

    Sun, X., Wu, J., Zhang, X., Zhang, Z., Zhang, C., Xue, T., Tenenbaum, J.B., Free- man, W.T.: Pix3d: Dataset and methods for single-image 3d shape modeling. In: Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2974–2983 (2018) 10, 23

    2018

  62. [62]

    In: Proc

    Thai, A., Stojanov, S., Upadhya, V., Rehg, J.M.: 3d reconstruction of novel object shapes from single images. In: Proc. of the International Conference on 3D Vision. pp. 85–95 (2021) 3

    2021

  63. [63]

    In: Proc

    Wang, N., Zhang, Y., Li, Z., Fu, Y., Liu, W., Jiang, Y.G.: Pixel2mesh: Generating 3d mesh models from single rgb images. In: Proc. of the European Conference on Computer Vision. pp. 52–67 (2018) 2, 3

    2018

  64. [64]

    In: Proc

    Wang, Z., Wang, Y., Chen, Y., Xiang, C., Chen, S., Yu, D., Li, C., Su, H., Zhu, J.: Crm: Single image to 3d textured mesh with convolutional reconstruction model. In: Proc. of the European Conference on Computer Vision. pp. 57–74 (2024) 2

    2024

  65. [65]

    In: Proc

    Wu, W., Qi, Z., Fuxin, L.: Pointconv: Deep convolutional networks on 3d point clouds. In: Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9621–9630 (2019) 2

    2019

  66. [66]

    In: Proc

    Xu, H., Lei, Y., Chen, Z., Zhang, X., Zhao, Y., Wang, Y., Tu, Z.: Bayesian diffusion models for 3d shape reconstruction. In: Proc. of the IEEE/CVF Conference on Point-MF 19 Computer Vision and Pattern Recognition. pp. 10628–10638 (2024) 2, 3, 10, 11, 12, 23

    2024

  67. [67]

    In: Proc

    Xu, Q., Xu, Z., Philip, J., Bi, S., Shu, Z., Sunkavalli, K., Neumann, U.: Point- nerf: Point-based neural radiance fields. In: Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5438–5448 (2022) 1

    2022

  68. [68]

    arXiv:2106.12102 (2021) 12

    Yagubbayli, F., Wang, Y., Tonioni, A., Tombari, F.: Legoformer: Transformers for block-by-block multi-view 3d reconstruction. arXiv:2106.12102 (2021) 12

    work page arXiv 2021

  69. [69]

    In: Advances in Neural Information Processing Systems

    Zeng, X., Vahdat, A., Williams, F., Gojcic, Z., Litany, O., Fidler, S., Kreis, K.: Lion: latent point diffusion models for 3d shape generation. In: Advances in Neural Information Processing Systems. pp. 10021–10039 (2022) 2, 3

    2022

  70. [70]

    In: Proc

    Zhang, H., Siarohin, A., Menapace, W., Vasilkovsky, M., Tulyakov, S., Qu, Q., Skorokhodov, I.: Alphaflow: Understanding and improving meanflow models. In: Proc. of the International Conference on Learning Representations (2026) 4

    2026

  71. [71]

    In: Proc

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 586–595 (2018) 4

    2018

  72. [72]

    In: Proc

    Zhou, C., Zhong, F., Hanji, P., Guo, Z., Fogarty, K., Sztrajman, A., Gao, H., Oztireli, C.: Frepolad: Frequency-rectified point latent diffusion for point cloud generation. In: Proc. of the European Conference on Computer Vision. pp. 434– 453 (2024) 2

    2024

  73. [73]

    In: Proc

    Zhou, L., Ermon, S., Song, J.: Inductive moment matching. In: Proc. of the Inter- national Conference on Machine Learning (2025) 4

    2025

  74. [74]

    In: Proc

    Zhu, D., Di, Y., Gavranovic, S., Ilic, S.: Sealion: Semantic part-aware latent point diffusion models for 3d generation. In: Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11789–11798 (2025) 2 20 Y. Baba and K. Yanai Supplementary Material A CFG-Guided Mean Flow Derivation of the Training Objective.To strengthen generatio...

    2025