Pith. sign in

REVIEW 3 major objections 5 minor 45 references

SeaLion: Semantic Part-Aware Latent Point Diffusion Models for 3D Generation

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read SeaLion claims that jointly predicting noise and part-segmentation labels inside a latent point diffusion model, with label-conditioned decoding, produces point clouds whose geometry and semantic part labels are both accurate and…

desk verdict Solid, honest extension of latent point diffusion to labeled generation; the new p-CD metric is useful but under-validated, and the SOTA claim needs more baselines and error bars. read the letter →

arxiv 2505.17721 v2 pith:ZPZBDV5R submitted 2025-05-23 cs.CV

Dekai Zhu , Yan Di , Stefan Gavranovic , Slobodan Ilic This is my paper · ORCID
classification cs.CV
keywords pointcloudgenerationdiffusionmodelssemanticpartsegmentationlatentpart-awareChamferdistance3Dshapeeditingsemi-supervisedlearninggenerativedataaugmentation
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

SeaLion asks whether a single diffusion model can generate 3D point clouds that come with correct per-point part labels, instead of generating an unlabeled cloud and labeling it afterwards. The paper's answer is yes: the point-level diffusion module is trained to predict, from the same intermediate features, both the noise added to latent points and the part-segmentation label of every point, and the decoder that turns latent points back into coordinates is conditioned on those labels. Because existing evaluation metrics such as 1-NNA ignore segmentation, the paper introduces part-aware Chamfer distance (p-CD), which sums per-part distances and treats point clouds with different part sets as infinitely far apart. SeaLion is reported to outperform the previous label-capable method DiffFacto by 13.33% and 6.52% on 1-NNA (p-CD) on ShapeNet and IntrA, and the authors show the model also supports semi-supervised training, generative data augmentation, and part-aware shape editing.

What carries the argument

The central machinery is the semantic part-aware latent point diffusion module $\epsilon_h$: a modified Point-Voxel CNN with one down-sampling path extracting shared representations and two parallel up-sampling paths that respectively predict the noise $\hat{\epsilon}_t$ and the segmentation encoding $\hat{y}_t$ for perturbed latent points $h_t$ at each diffusion step. The companion mechanism is part-aware Chamfer distance (p-CD), defined as the sum over parts of the bidirectional Chamfer distance between corresponding parts, with infinite distance when the part sets differ; feeding p-CD into 1-NNA is what lets the evaluation reward shapes whose parts are individually faithful and coherently joined. The diffusion module carries the generation and label prediction; p-CD carries the claim that a low score certifies both geometry and segmentation rationality.

What would settle it

Take SeaLion-generated point clouds, randomly permute their part labels, and recompute 1-NNA (p-CD) against the real test set; if the permuted clouds score nearly the same as the originals, p-CD is insensitive to label correctness and the segmentation claim is unsupported.

Watch

Extended reading notes

Core claim

The central discovery, as the paper states it, is that semantic part awareness can be built directly into latent-point diffusion: rather than factorizing generation part by part as DiffFacto does, SeaLion diffuses all latent points jointly and uses a shared down-sampling path with two parallel up-sampling branches so that one network emits both the predicted noise and the predicted segmentation labels at every denoising step. At inference, an exponential moving average over label predictions conditions the decoder, strengthening the alignment between coordinates and labels. The authors claim this design avoids the part-to-part incoherence of DiffFacto, yields state-of-the-art 1-NNA (p-CD), COV (p-CD), and MMD (p-CD) on ShapeNet and IntrA, and does not hurt unlabeled generation quality relative to Lion when the segmentation branch is ignored. They further claim that low p-CD between a generated cloud and a real cloud implies both geometric quality and reasonable part structure, because p-CD only vanishes when every generated part aligns with the corresponding real part.

Load-bearing premise

The result hinges on p-CD being a valid proxy for label correctness: if a generated cloud whose parts are geometrically plausible but mislabeled could still achieve a low p-CD to a real cloud, then the headline 1-NNA improvements would not prove that SeaLion generates accurate segmentation labels.

Editorial extensions

If this is right

  • Generated labeled point clouds no longer require a separate pretrained segmentation model to assign pseudo-labels.
  • Part-aware editing becomes a latent-space operation: freeze the latent points of a chosen part and diffuse the remaining points.
  • Generated samples can enlarge labeled training sets for downstream point-cloud segmentation models, including in domains with scarce labels such as medical imaging.
  • Semi-supervised training works with the same architecture by zeroing the label conditioning and dropping the segmentation loss term on unlabeled samples.
  • Evaluation metrics that previously ignored labels can be upgraded simply by replacing Chamfer distance with p-CD.

Reading between the lines

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

  • p-CD could plausibly be used as a training-time objective or editing guidance signal, not only as a post-hoc metric, since it directly rewards coherent part alignment.
  • The two-branch design suggests a general recipe: any latent point diffusion model can gain a discriminative label head with little architectural change, without degrading generation quality.
  • The paper's implicit equation of low p-CD with correct labels is worth stress-testing on fine-grained part sets where geometric alignment and label identity can diverge.
  • If the semi-supervised result extends, the same model could be pushed toward discovering parts with no labels at all, with segmentation annotations applied only at inference.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper presents SeaLion, a latent point diffusion model that generates 3D point clouds with per-point semantic part labels. The method builds on Lion's hierarchical latent diffusion: a point-level encoder/decoder is conditioned on segmentation encodings, and the point-level diffusion module is trained to jointly predict the noise and the segmentation labels from the same down-sampled representation. At inference, the predicted labels are smoothed by EMA and used to condition the decoder, producing a labeled point cloud. The paper also proposes a pairwise distance, part-aware Chamfer distance (p-CD), and redefines 1-NNA, COV, and MMD using p-CD. Experiments on ShapeNet and IntrA compare SeaLion with DiffFacto and with two-step baselines (Lion plus a pretrained segmenter), and additional experiments address semi-supervised training, generative data augmentation for segmentation, and part-aware editing.

Significance. If the results hold, SeaLion offers a practical single-stage alternative to generating semantically labeled point clouds without an external post-hoc segmenter, and p-CD is a natural extension of Chamfer distance to labeled clouds. The paper is strong on reproducibility: architecture hyperparameters are tabulated in the supplementary, code is released, and the ablation in Table 7 explicitly checks that the auxiliary segmentation branch does not degrade unlabeled generation quality. The downstream data-augmentation experiment (Table 5) provides an independent, practically meaningful check that the generated labels are useful. The main weakness is that the headline claim of generating 'accurate segmentation labels' is supported primarily by p-CD-based metrics, which are proposed by the authors and not directly validated against any reference label field on generated shapes; the comparison to DiffFacto also rests on a single baseline and on four ShapeNet categories. These issues are addressable but require additional evaluation before the central claim is fully established.

major comments (3)
  1. [§3.4, §A.2] The paper's central claim that SeaLion generates point clouds with accurate part labels is supported mainly by 1-NNA (p-CD), whose validity as a proxy for label correctness rests on Argument 3 in Supplementary A.2. Argument 3 asserts that a small per-part Chamfer distance to a real cloud implies that the generated cloud is segmented well, but this is not empirically validated. p-CD is a per-part geometric distance, so it does not directly measure semantic correctness of the label field in cases where part geometry is ambiguous, part boundaries are slightly shifted, or a generated part is broken into multiple components that still have small per-part distance to a single real part. The paper provides no direct comparison of the final predicted label field y_hat against any reference on generated shapes, and Figure 5 reports mIoU only for intermediate diffusion steps on a single class. Please add a direct label-accuracy evaluation: for real point clouds, compute the accuracy/mIoU of y_hat against ground-truth labels (extending Figure 5 to all categories and to the final smoothed prediction); for generated clouds, report agreement between the output labels and labels assigned by a strong pretrained segmenter on the same geometry, or a label-consistency metric. This would directly support the 'accurate segmentation labels' claim and also validate p-CD as a proxy.
  2. [§4.2, Table 1] The headline state-of-the-art claim ('outperforming the existing state-of-the-art model, DiffFacto, by 13.33% and 6.52% on 1-NNA (p-CD)') is based on a single baseline and on only four ShapeNet categories, and the paper reports no variance or significance information. Since 1-NNA is computed on a finite set of generated samples, the observed difference could be within sampling noise, especially for categories with small test sets. Please report standard deviations over multiple generation runs or bootstrap confidence intervals, and either add additional labeled-generation baselines (e.g., a VAE/GAN-based part-label generator) or temper the claim to 'outperforms the compared baselines'.
  3. [§4.3, Table 4] The semi-supervised experiment shows that adding unlabeled data improves over using 10% labels alone, but it does not include a fully supervised upper bound (SeaLion trained with 100% labeled data). Without this bound, the claim that SeaLion 'reduces the demand for labeling efforts' is not quantified: the reader cannot tell whether the semi-supervised model approaches fully supervised performance or remains far below it. Please add the 100%-labeled result for the same car category.
minor comments (5)
  1. [§3.1, Eq. (6), Eq. (8)] The values of the balancing hyperparameters λseg, λz, and λh are not given in the paper or in the supplementary material; please provide them for reproducibility.
  2. [§4.2, Table 1] Please state explicitly how many point clouds are generated per category for evaluation, since 1-NNA and COV are sensitive to the size of the generated set G.
  3. [Figure 5] The caption says 'Evolution of predictive performance measured by mIoU for different diffusion steps t on airplane class,' but it does not specify the dataset split (train or test) or whether the mIoU is computed between the predicted y_t and ground-truth labels of real point clouds; please clarify.
  4. [§3.4] The statement that p-CD is infinite for point clouds with different part sets could make 1-NNA degenerate if a generated cloud is missing a part; please discuss how this case is handled in the evaluation and whether any generated samples were excluded for this reason.
  5. [References] The paper cites ShapeNet as [37] (Yi et al.) for the segmentation-labeled subset and [3] (Chang et al.) for the larger ShapeNet repository; consider adding a sentence clarifying the relationship between these two references to avoid reader confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the generation and evaluation claims are self-contained, and the p-CD metric is not fitted to the model.

full rationale

The paper's central derivation chain consists of the SeaLion training objective (Eqs. 6–8) and the p-CD pairwise distance (Eq. 11). These do not reduce to each other. The model is trained with a VAE ELBO, a diffusion noise-prediction loss, and a cross-entropy segmentation term; p-CD is never used as a training loss, and no parameter is fitted to it. The headline SOTA numbers are obtained by applying p-CD to held-out generated sets and real test sets, which is a post-hoc evaluation rather than a construction. The main same-author element is the introduction of p-CD itself, but the paper defines it explicitly as a sum of per-part Chamfer distances and provides an external argument for its behavior in Suppl. A.2. Introducing a metric is not circular unless the metric is used as a training objective or its definition already encodes the target result; neither is the case here. The downstream data augmentation experiment (Table 5) provides an independent check that the generated labels are useful, since improved mIoU on real test data cannot be manufactured by the metric alone. No equation in the paper is equivalent by construction to another, and no uniqueness theorem is imported from the authors' prior work. The concern that p-CD may not fully verify semantic label accuracy is a validity or correctness issue, not a circularity reduction. Therefore the paper is self-contained and receives a circularity score of 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard diffusion math, on the adopted Lion framework, on the cited assumption that diffusion features carry semantic information, and on the paper's own assumption that p-CD tracks label correctness. Free parameters are standard training hyperparameters; no invented entities are introduced.

free parameters (4)
  • λseg (segmentation loss weight)
    Weights the cross-entropy term H(y, ŷ_t) in Eq. (8); chosen by hand and affects the balance between noise prediction and label prediction.
  • λz and λh (KL divergence weights)
    Balance reconstruction accuracy and latent regularization in Eq. (6); chosen by hand and affect the learned latent space.
  • EMA smoothing factor α = 0.1
    Refines predicted labels y_t over denoising steps during inference (Section 3.1); chosen by hand and affects final segmentation quality.
  • Edit perturbation steps τ
    Number of steps for perturbing unfrozen latent points in the editing algorithm (Section 3.3, Algorithm 1); chosen by the user, controls edit strength.
assumptions (5)
  • standard math DDPM training and sampling procedures are valid as described by Ho et al. 2020.
    Eqs. (1), (7), and (8) use the standard DDPM objective without re-derivation.
  • domain assumption Lion's latent point diffusion framework (global latent z0 and point-latents h0) provides an effective representation for 3D generation.
    Section 3.1 adopts Lion's hierarchy; if Lion's representation was ineffective, SeaLion's design would inherit the failure.
  • domain assumption Intermediate features of diffusion models capture semantic information that transfers to segmentation (Baranchuk et al. 2021).
    Motivates jointly predicting noise and labels from the same network (Eq. (3)); cited but not proven in this paper.
  • ad hoc to paper p-CD is a valid proxy for both geometric quality and semantic label correctness.
    Supplementary A.2 Argument 3 claims small p-CD to a real sample implies 'xg is segmented well'; this assumption underpins the main evaluation.
  • domain assumption The decoder conditioned on predicted labels y improves alignment between geometry and labels.
    Inference in Section 3.1 uses ŷ to decode latent points; relies on the joint predictions being accurate enough that mispredictions do not degrade geometry.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SeaLion: Semantic Part-Aware Latent Point Diffusion Models for 3D Generation." pith.science (2026). https://pith.science/paper/ZPZBDV5R

@misc{pith2026250517721,
  author       = {Pith},
  title        = {Pith review of: SeaLion: Semantic Part-Aware Latent Point Diffusion Models for 3D Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZPZBDV5R}},
  note         = {Machine review of arXiv:2505.17721}
}
read the original abstract

Denoising diffusion probabilistic models have achieved significant success in point cloud generation, enabling numerous downstream applications, such as generative data augmentation and 3D model editing. However, little attention has been given to generating point clouds with point-wise segmentation labels, as well as to developing evaluation metrics for this task. Therefore, in this paper, we present SeaLion, a novel diffusion model designed to generate high-quality and diverse point clouds with fine-grained segmentation labels. Specifically, we introduce the semantic part-aware latent point diffusion technique, which leverages the intermediate features of the generative models to jointly predict the noise for perturbed latent points and associated part segmentation labels during the denoising process, and subsequently decodes the latent points to point clouds conditioned on part segmentation labels. To effectively evaluate the quality of generated point clouds, we introduce a novel point cloud pairwise distance calculation method named part-aware Chamfer distance (p-CD). This method enables existing metrics, such as 1-NNA, to measure both the local structural quality and inter-part coherence of generated point clouds. Experiments on the large-scale synthetic dataset ShapeNet and real-world medical dataset IntrA demonstrate that SeaLion achieves remarkable performance in generation quality and diversity, outperforming the existing state-of-the-art model, DiffFacto, by 13.33% and 6.52% on 1-NNA (p-CD) across the two datasets. Experimental analysis shows that SeaLion can be trained semi-supervised, thereby reducing the demand for labeling efforts. Lastly, we validate the applicability of SeaLion in generative data augmentation for training segmentation models and the capability of SeaLion to serve as a tool for part-aware 3D shape editing.

Figures

Figures reproduced from arXiv: 2505.17721 by the authors.

Figure 1
Figure 1. Leveraging the proposed semantic part-aware latent [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. (a) Training: The generative model develops semantic part awareness by being trained to reconstruct input point clouds x guided by segmentation encodings y, and to jointly predict the noise ϵˆt for perturbed latent points ht and segmentation labels yˆt at diffusion step t. (b) Inference: Starting from Gaussian noise, the diffusion modules generate z0, h0, and yˆ. Then, the conditional decoding guided by z0 and yˆ ge… view at source ↗
Figure 3
Figure 3. Data flow in the point-level diffusion module [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Limitations of the intra-part (1-NNA-P) and inter-part [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 6
Figure 6. Figure 6: Up: Generated point clouds of airplanes, cars, chairs, guitars, lamps, and tables from SeaLion. Bottom: Generated aneurysm segments from SeaLion (red: vessels, blue: aneurysm). 6 [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 5
Figure 5. Figure 5: Evolution of predictive perfor￾mance measured by mIoU for different diffu￾sion steps t on airplane class. The prediction accuracy improves as t decreases from T to 0. lamp categories, SeaLion outperforms DiffFacto by an av￾erage of 13.33% on 1-NNA (p-CD), 11.61% on COV…
Figure 7
Figure 7. Figure 7: (a) Original point clouds and (b) novel generated point [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: The generated set G, which either (a) exhibits poor mode coverage compared to the real dataset R or (b) contains poor-quality samples, cannot achieve a good 1-NNA score. The arrows indicate the nearest neighbors of samples. In both cases, most samples and their nearest…
Figure 9
Figure 9. Figure 9: Generated point clouds of airplane class from SeaLion. [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: Generated point clouds of car class from SeaLion. [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Generated point clouds of chair class from SeaLion. [PITH_FULL_IMAGE:figures/full_fig_p016_11.png]
Figure 12
Figure 12. Figure 12: Generated point clouds of guitar class from SeaLion. [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]
Figure 13
Figure 13. Figure 13: Generated point clouds of lamp class from SeaLion. [PITH_FULL_IMAGE:figures/full_fig_p017_13.png]
Figure 14
Figure 14. Figure 14: Generated point clouds of table class from SeaLion. [PITH_FULL_IMAGE:figures/full_fig_p017_14.png]
Figure 15
Figure 15. Figure 15: Comparison of cars generated by SeaLion, Lion & SPoTr [ [PITH_FULL_IMAGE:figures/full_fig_p017_15.png]
Figure 16
Figure 16. Figure 16: Examples of implausible inter-part coherence in shapes generated by DiffFacto [ [PITH_FULL_IMAGE:figures/full_fig_p018_16.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 35 canonical work pages

  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. 2, 5, 11

    work page 2018

  2. [2]

    Label-efficient se- mantic segmentation with diffusion models

    Dmitry Baranchuk, Ivan Rubachev, Andrey V oynov, Valentin Khrulkov, and Artem Babenko. Label-efficient se- mantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. 2, 3, 7

    arXiv 2021

  3. [3]

    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. 13

    arXiv 2015

  4. [4]

    Generative pre- training from pixels

    Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Hee- woo Jun, David Luan, and Ilya Sutskever. Generative pre- training from pixels. In International conference on machine learning, pages 1691–1703. PMLR, 2020. 3

    work page 2020

  5. [5]

    A comprehensive survey for generative data augmentation

    Yunhao Chen, Zihui Yan, and Yunjie Zhu. A comprehensive survey for generative data augmentation. Neurocomputing, page 128167, 2024. 8

    work page 2024

  6. [6]

    Gen- erative adversarial networks: An overview

    Antonia Creswell, Tom White, Vincent Dumoulin, Kai Arulkumaran, Biswa Sengupta, and Anil A Bharath. Gen- erative adversarial networks: An overview. IEEE signal pro- cessing magazine, 35(1):53–65, 2018. 1

    work page 2018

  7. [7]

    Ccd-3dr: Consistent conditioning in diffusion for single-image 3d reconstruction

    Yan Di, Chenyangguang Zhang, Pengyuan Wang, Guangyao Zhai, Ruida Zhang, Fabian Manhardt, Benjamin Busam, Xi- angyang Ji, and Federico Tombari. Ccd-3dr: Consistent conditioning in diffusion for single-image 3d reconstruction. arXiv preprint arXiv:2308.07837, 2023. 1

    arXiv 2023

  8. [8]

    Large scale adversar- ial representation learning

    Jeff Donahue and Karen Simonyan. Large scale adversar- ial representation learning. Advances in neural information processing systems, 32, 2019. 3

    work page 2019

Show all 45 references
  1. [9]

    Ad- versarial feature learning

    Jeff Donahue, Philipp Kr ¨ahenb¨uhl, and Trevor Darrell. Ad- versarial feature learning. arXiv preprint arXiv:1605.09782,

    arXiv

  2. [10]

    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. 1

    2017

  3. [11]

    Diffusion mod- els and representation learning: A survey

    Michael Fuest, Pingchuan Ma, Ming Gui, Johannes S Fis- cher, Vincent Tao Hu, and Bjorn Ommer. Diffusion mod- els and representation learning: A survey. arXiv preprint arXiv:2407.00783, 2024. 12

    2024 arXiv

  4. [12]

    Mrgan: Multi-rooted 3d shape representation learning with unsupervised part disentanglement

    Rinon Gal, Amit Bermano, Hao Zhang, and Daniel Cohen- Or. Mrgan: Multi-rooted 3d shape representation learning with unsupervised part disentanglement. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 2039–2048, 2021. 1, 2

    2021

  5. [13]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 5

    2016

  6. [14]

    Spaghetti: Editing implicit shapes through part aware generation.ACM Transactions on Graph- ics (TOG), 41(4):1–20, 2022

    Amir Hertz, Or Perel, Raja Giryes, Olga Sorkine-Hornung, and Daniel Cohen-Or. Spaghetti: Editing implicit shapes through part aware generation.ACM Transactions on Graph- ics (TOG), 41(4):1–20, 2022. 2

    2022

  7. [15]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020. 1, 3

    2020

  8. [16]

    Pf-net: Point fractal network for 3d point cloud completion

    Zitian Huang, Yikuan Yu, Jiawen Xu, Feng Ni, and Xinyi Le. Pf-net: Point fractal network for 3d point cloud completion. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 7662–7670, 2020. 1

    2020

  9. [17]

    Gal: Geometric adversarial loss for single-view 3d-object recon- struction

    Li Jiang, Shaoshuai Shi, Xiaojuan Qi, and Jiaya Jia. Gal: Geometric adversarial loss for single-view 3d-object recon- struction. In Proceedings of the European conference on computer vision (ECCV), pages 802–816, 2018. 1

    2018

  10. [18]

    Auto-encoding varia- tional bayes

    Diederik P Kingma and Max Welling. Auto-encoding varia- tional bayes. arXiv preprint arXiv:1312.6114, 2013. 1

    2013 arXiv

  11. [19]

    Your diffusion model is secretly a zero-shot classifier

    Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 2206–2217,

  12. [20]

    Sp- gan: Sphere-guided 3d shape generation and manipulation

    Ruihui Li, Xianzhi Li, Ka-Hei Hui, and Chi-Wing Fu. Sp- gan: Sphere-guided 3d shape generation and manipulation. ACM Transactions on Graphics (TOG), 40(4):1–12, 2021. 2

    2021

  13. [21]

    Editvae: Unsupervised parts-aware controllable 3d point cloud shape generation

    Shidi Li, Miaomiao Liu, and Christian Walder. Editvae: Unsupervised parts-aware controllable 3d point cloud shape generation. In Proceedings of the AAAI Conference on Arti- ficial Intelligence, pages 1386–1394, 2022. 1, 2

    2022

  14. [22]

    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. Advances in Neural Information Processing Systems, 32, 2019. 4, 14

    2019

  15. [23]

    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. 1, 2

    2021

  16. [24]

    SDEdit: Guided image synthesis and editing with stochastic differential equa- tions

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia- jun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equa- tions. In International Conference on Learning Representa- tions, 2022. 5

    2022

  17. [25]

    Difffacto: Controllable part-based 3d point cloud generation with cross diffusion

    George Kiyohiro Nakayama, Mikaela Angelina Uy, Jiahui Huang, Shi-Min Hu, Ke Li, and Leonidas Guibas. Difffacto: Controllable part-based 3d point cloud generation with cross diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14257–14267...

    2023

  18. [26]

    3d part segmentation on shapenet-part

    Papers with Code. 3d part segmentation on shapenet-part. https://paperswithcode.com/sota/3d-part- segmentation - on - shapenet - part, 2024. Ac- cessed: 2024-11-14. 6

    2024

  19. [27]

    Self-positioning point-based transformer for point cloud understanding

    Jinyoung Park, Sanghyeok Lee, Sihyeon Kim, Yunyang Xiong, and Hyunwoo J Kim. Self-positioning point-based transformer for point cloud understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21814–21823, 2023. 6, 8, 12, 14, 17

    2023

  20. [28]

    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. 4, 6, 12, 14 9

    2017

  21. [29]

    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. In- ternational journal of computer vision, 40:99–121, 2000. 2, 5

    2000

  22. [30]

    3d point cloud generative adversarial network based on tree structured graph convolutions

    Dong Wook Shu, Sung Woo Park, and Junseok Kwon. 3d point cloud generative adversarial network based on tree structured graph convolutions. In Proceedings of the IEEE/CVF international conference on computer vision , pages 3859–3868, 2019. 1, 2

    2019

  23. [31]

    Repurposing gans for one-shot se- mantic part segmentation

    Nontawat Tritrong, Pitchaporn Rewatbowornwong, and Su- pasorn Suwajanakorn. Repurposing gans for one-shot se- mantic part segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 4475–4485, 2021. 3

    2021

  24. [32]

    Linear semantics in genera- tive adversarial networks

    Jianjin Xu and Changxi Zheng. Linear semantics in genera- tive adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 9351–9360, 2021

    2021

  25. [33]

    Generative hierarchical features from synthe- sizing images

    Yinghao Xu, Yujun Shen, Jiapeng Zhu, Ceyuan Yang, and Bolei Zhou. Generative hierarchical features from synthe- sizing images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 4432– 4442, 2021. 3

    2021

  26. [34]

    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 Proceed- ings of the IEEE/CVF international conference on computer vision, pages 4541–4550, 2019. 2, 5, 6, 11

    2019

  27. [35]

    Dsg-net: Learning disentangled structure and geometry for 3d shape generation

    Jie Yang, Kaichun Mo, Yu-Kun Lai, Leonidas J Guibas, and Lin Gao. Dsg-net: Learning disentangled structure and geometry for 3d shape generation. ACM Transactions on Graphics (TOG), 42(1):1–17, 2022. 2

    2022

  28. [36]

    Intra: 3d intracranial aneurysm dataset for deep learning

    Xi Yang, Ding Xia, Taichi Kin, and Takeo Igarashi. Intra: 3d intracranial aneurysm dataset for deep learning. In CVPR, pages 2656–2666, 2020. 2, 6, 7, 12, 13

    2020

  29. [37]

    A scalable active framework for region annotation in 3d shape collections.ACM Transactions on Graphics (ToG), 35(6):1–12, 2016

    Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Shef- fer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections.ACM Transactions on Graphics (ToG), 35(6):1–12, 2016. 2, 5, 6, 7, 8, 12, 13, 14

    2016

  30. [38]

    Pointr: Diverse point cloud comple- tion with geometry-aware transformers

    Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud comple- tion with geometry-aware transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12498–12507, 2021. 1

    2021

  31. [39]

    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 (NeurIPS), 2022. 1, 2, 3, 4, 5, 6, 7, 12, 13, 14, 17

    2022

  32. [40]

    Point cloud part editing: Segmentation, generation, assembly, and selection

    Kaiyi Zhang, Yang Chen, Ximing Yang, Weizhong Zhang, and Cheng Jin. Point cloud part editing: Segmentation, generation, assembly, and selection. In Proceedings of the AAAI Conference on Artificial Intelligence , pages 7187– 7195, 2024. 1

    2024

  33. [41]

    Datasetgan: Efficient labeled data factory with minimal human effort

    Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean- Francois Lafleche, Adela Barriuso, Antonio Torralba, and Sanja Fidler. Datasetgan: Efficient labeled data factory with minimal human effort. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognitio...

    2021

  34. [42]

    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 Proceed- ings of the IEEE/CVF International Conference on Com- puter Vision, pages 5826–5835, 2021. 1, 2, 5 10 SeaLion: Semantic Part-Aware Latent Point Diffusion Models f...

    2021

  35. [43]

    Semantic Part-Aware Latent Point Diffusion

    Methodology 3 3.1. Semantic Part-Aware Latent Point Diffusion . 3 3.2. Model Architecture of SeaLion . . . . . . . 4 3.3. Part-aware 3D Shape Edition Tool . . . . . . 5 3.4. Evaluation Metrics . . . . . . . . . . . . . . 5

  36. [44]

    Experimental Setup

    Experiments 6 4.1. Experimental Setup . . . . . . . . . . . . . . 6 4.2. Experimental Results . . . . . . . . . . . . . 6 4.3. Experimental Analysis . . . . . . . . . . . . 7 4.4. Applications . . . . . . . . . . . . . . . . . 8

  37. [45]

    ground truth

    Conclusion 8 A . Evaluation Metrics 11 A.1 . Calculation Formulas . . . . . . . . . . . . . 11 A.2 . More Discussions about Part-aware Metrics . 11 B . Pseudo-code of Part-aware 3D Editing 12 C . Additional Experimental Details and Results 12 C.1. Two-step Method on IntrA Data...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.