Learning to Tessellate: Point Cloud Generation via Recursive Spectral Partitioning

Bangzhen Liu; Huaidong Zhang; Monan Sun; Shengfeng He

REVIEW 5 major objections 5 minor 42 references

Learning to Tessellate: Point Cloud Generation via Recursive Spectral Partitioning

T0 review · 5 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read PointRSP claims that replacing heuristic token orderings with recursive spectral partitioning into a non-balanced binary tree yields state-of-the-art autoregressive point cloud generation, besting diffusion baselines on ShapeNet benchmarks.

desk verdict A genuinely new tokenization idea—recursive spectral partitioning into a non-balanced tree—but the SOTA claim is under-supported by a re-instantiated baseline and missing error bars. read the letter →

arxiv 2608.02432 v1 pith:BEJ4DSPS submitted 2026-08-03 cs.CV

Monan Sun , Bangzhen Liu , Huaidong Zhang , Shengfeng He This is my paper

classification cs.CV

keywords pointcloudgenerationautoregressivemodelrecursivespectralpartitioningFiedlervectortopology-preservingtessellationnon-balancedbinarytreedual-streamcascadedgeneratorShapeNetV2

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

The paper argues that the ordering of tokens in autoregressive point cloud generation should come from the shape's own geometry rather than from a heuristic convention. To that end it proposes PointRSP, which recursively partitions a point cloud into a non-balanced binary tree using spectral graph cuts, then generates the tree level-by-level with a dual-stream transformer that predicts both structure and residual features. On the standard ShapeNetV2 single-category benchmarks the method reports the best 1-NNA accuracy among all compared diffusion and autoregressive models, under both Chamfer Distance and Earth Mover's Distance. The authors take this as evidence that topology-aware tokenization is the decisive factor, not the generator backbone or the diffusion-versus-autoregressive paradigm.

What carries the argument

The recursive spectral bipartitioner: at each level, the sign of the Fiedler vector (the eigenvector of the graph Laplacian for its second-smallest eigenvalue) of a KNN graph over token features decides which side of a split each token belongs to. Combined with a bottom-up greedy merge stage, it creates a non-balanced binary tree whose variable-depth leaves accommodate natural asymmetry in 3D shapes. This tree is the load-bearing object: it defines the structural labels for quantization, the ordering of the autoregressive sequence, and the anchor centers for position encoding.

What would settle it

Take a set of shapes with semantic part annotations (e.g., airplane wings/fuselage/tail, chair legs/seat/back), run PointRSP's partitioning autoencoder, and measure how often a spectral cut separates two points belonging to the same part or merges points from different parts. If a substantial fraction of cuts violate part boundaries, the claimed topology preservation is not happening; alternatively, compare persistent homology of generated vs. real point clouds and check whether the gap over heuristic-order baselines vanishes.

Watch

Extended reading notes

Core claim

The paper's central discovery is that a shape's token sequence for autoregressive generation should be derived from its own geometry, not from a global convention. PointRSP decomposes a point cloud into a non-balanced binary tree: bottom-up greedy merging groups local details into K tokens, then top-down spectral bisection splits those tokens using the sign of the Fiedler vector of a nearest-neighbour graph Laplacian at each recursion. The resulting tree is deterministic and per-shape, and it is used both to quantize pooled features into a codebook and to guide a dual-stream cascaded transformer that predicts branching structure and residual feature synthesis. The authors report consistent s

Load-bearing premise

The sign of the Fiedler vector of a nearest-neighbor graph Laplacian computed on pooled features is assumed to give a stable, geometrically meaningful bisection of every object at every recursion depth; if any cut tears across a thin structure or noisy feature pool, the 'topology-preserving blueprint' becomes an arbitrary partition and the generator is conditioning on noise.

Editorial extensions

If this is right

Heuristic tokenizations (coordinate sorting, space-filling curves, stochastic downsampling) are replaced by a deterministic per-shape spectral tree; generation quality no longer depends on hand-chosen orderings.
The non-balanced binary tree overcomes the strict 2^t label constraint of balanced hierarchical quantizers, allowing deeper splits where geometry is complex and shallower splits elsewhere.
Anchoring early tokens to multiscale structural centers stabilizes the cold-start phase of cascaded generation, as shown by reduced batch-level MMD variance in ablations.
The same topology-aware blueprint improves class-conditional generation and serves as a point cloud upsampling module, indicating it is a transferable representation, not just a generator detail.

Reading between the lines

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

A direct test of the paper's 'topology-preserving' claim would compare the learned tree against human part hierarchies (e.g., wings vs fuselage) or persistent-homology descriptors; the paper itself does not provide such a metric.
Because the Fiedler vector is just one cut criterion, one could try multiway spectral cuts or thresholding at eigenvector gaps; the paper's hybrid merge-then-split design suggests any such variant could be swapped in without changing the generator.
If the blueprint is truly deterministic and category-agnostic, it could serve as a shared tokenizer for shape completion, editing, or single-view reconstruction—tasks the paper does not address.
The reported gains are on object-centric shapes at 2,048 points; the paper's own limitations note that ultra-dense point sets and scene-level data are untested, so the 'strong generalization across complex topologies' claim is provisional.

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.

Desk Editor's Note

read the letter

The one thing worth knowing about this paper: the tokenizer is genuinely new. PointRSP replaces heuristic orderings (sorting, space-filling curves, stochastic subsets) with a deterministic, non-balanced binary tree built by bottom-up greedy merging followed by top-down spectral bisection. That is a real departure from the field's default playbook, and it's a plausible way to give an autoregressive generator a stable structural prior. The dual-stream cascaded generator and the geometry-calibrated positional encoding are sensible engineering around the cold-start problem, and the ablation in Table 3 gives internal evidence that each component earns its keep. The paper is also honest about its scope—object-centric shapes, standard densities—which I appreciate.

The soft spot is the experimental case for "state-of-the-art." The headline numbers in Table 1 depend on a PointNSP baseline that the authors re-instantiated themselves, using its VAE module with a different generator, rather than running the official code or numbers. That's not automatically fatal, but the text should say so more prominently, and the margins are thin enough (mean EMD 56.86 vs 57.08; chair CD 53.88 vs 55.15) that one standard error could flip the ranking. There are no error bars, no multiple seeds, no released configs. The stress-test note is right about this being the main vulnerability. The secondary concern about spectral cuts not always aligning with semantic parts is interesting but less important—a deterministic tree could still be a good conditioning signal even if it doesn't perfectly respect topology. Still, the paper's claim to be "topology-preserving" would be stronger with a direct measure of structural fidelity rather than only 1-NNA.

The equations are also sloppier than they should be: Eq. 3 overloads the index, Eq. 9 uses a sigmoid with >0 which is trivially true, and Eq. 10 is underspecified. These are minor and fixable, not fundamental.

Who gets value from this? Anyone working on autoregressive point cloud generation or hierarchical tokenization for 3D data. The core idea is worth engaging with and citing, even if the experimental evidence needs strengthening. I would send this to peer review—the novelty and internal coherence justify referee time. The revision should request official baselines, variance across seeds, and code. If those hold up, this could be a solid contribution rather than just a promising one.

Referee Report

5 major / 5 minor

Summary. The paper proposes PointRSP, an autoregressive point cloud generator that replaces heuristic token orderings with a deterministic hierarchical tokenization built from recursive spectral partitioning. A topology-aware partitioning autoencoder merges point tokens greedily and then recursively bisects the resulting nearest-neighbor graph using Fiedler-vector signs, producing a non-balanced binary tree; a dual-stream cascaded transformer predicts structural splits and residual features, with a geometry-calibrated positional encoding to stabilize early steps. Experiments on ShapeNetV2 report state-of-the-art 1-NNA for single-category generation (Table 1), plus multi-class and upsampling results (Table 2) and ablations (Table 3).

Significance. If the reported results are reproducible, the paper makes a solid contribution: it gives a principled, deterministic alternative to coordinate sorting and stochastic downsampling for AR point-cloud tokenization, and the ablation study (Table 3) supports the incremental value of the proposed components. The method also combines structural prediction and feature synthesis in a clean way. However, the central SOTA claim rests on a single unseeded run of a re-instantiated baseline, and several equations that define the proposed blueprint are underspecified. The contribution is therefore plausible but not yet fully verified.

major comments (5)

[§4.2, Table 1] The state-of-the-art claim is not fully supported as reported. §4.2 states that the PointNSP baseline was 'instantiated' by the authors using its quantization module with a different generator, not by running the official PointNSP implementation. Since PointNSP's generator and training schedule are part of its design, the re-implementation may be weaker. Margins are small (e.g., Chair EMD 53.96 vs 54.17, Car CD 51.47 vs 53.66). Please report official PointNSP numbers (or release code/configs and compare under identical training), and give 1-NNA with standard errors over at least three seeds.
[Eq. (9)] The structural prediction is written as σ(Attns(...)) > 0. Since a sigmoid is always positive, this inequality is always true and does not produce binary structural decisions. A threshold of 0.5 or Bernoulli sampling is needed. As written, the 'probabilistic binary decisions' are ill-defined, which directly affects the structural labels used in Eqs. (10)–(11) and the claimed binary-tree blueprint.
[Eq. (10)] The geometry-calibrated positional encoding is underspecified. The term P(map(i, eSt), j) is not defined: P_t is a set of group centers, but the formula needs P_j(map(i, eSt)) and a precise definition of historical scales. Also, the exponentially increasing weights w_{j+1}=2w_j with normalization Σ w_j=1 make the weights depend on t; this must be stated explicitly, or the encoding at different t is not comparable.
[Eqs. (3)–(5)] The blueprint indexing is confusing and internally inconsistent. Eq. (3) uses j both as the summation index and as an argument of s^(2)(i,j) on the left-hand side. Eqs. (4)–(5) mix indices t, l_k, T_l, T_k without a precise statement of the order in which merge and split scales are enumerated; since l_k varies per token, the composition s(2)(i,t) = ssplit(k,t) for t<l_k and smerge(i,t−l_k) for l_k≤t<l_k+T_l is not a single well-defined scale sequence. Please rewrite the indexing with explicit ranges.
[§1, Eq. (4), Table 3] The paper's central conceptual claim is 'topology preservation', but no quantity that measures topological fidelity is reported. The Fiedler-vector sign split (Eq. 4) is assumed to give a meaningful geometry-preserving bisection at every level; this is never validated (e.g., by stability of the split under KNN-graph perturbations, or by component/boundary metrics). Please add at least one topological diagnostic alongside the 1-NNA numbers, or explicitly soften the topology-preservation claim to 'structure-consistent'.

minor comments (5)

[§3.1] Typo: 'hybird' should be 'hybrid'.
[Eq. (6)] Loss weights λ_L1, λ_cmf, λ_vq are not given in the main text; if they are in the supplementary, the supplementary is not included.
[§4.3/Table 2] The sentence 'left forms of Table 2' should read 'left columns'; also the multi-class Chair EMD is worse than PointNSP (56.07 vs 55.10), so the text should explicitly say which two experiments are 'competitive' rather than SOTA.
[§4.4/Table 3] The ablation reports only MMD; adding 1-NNA with standard deviation would strengthen the claim that each component improves both quality and diversity.
[§2] The related-work discussion of visual AR models is brief, but acceptable; consider citing the original RQ-VAE when discussing residual quantization.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claim is an empirical SOTA result evaluated against external benchmarks and baselines.

full rationale

The paper's central claim is an empirical state-of-the-art result evaluated on ShapeNetV2 against external baselines (PVD, LION, TIGER, PointGPT, PointNSP). The recursive spectral partitioning in Eqs. 4–5 defines a deterministic, data-dependent structural labeling from the input point features; this labeling is then used as a conditioning/supervision target for the autoencoder and generator. This is standard tokenization/quantization training, not a prediction derived from the model's own outputs. The generation loss in Eq. 11 compares generated features/structures to the autoencoder's ground-truth representations, which is normal autoencoding/generative supervision. The only self-citations ([13] GenPoly, [41] Recdreamer) appear in related-work context and are not load-bearing for the proposed architecture. The re-instantiated PointNSP baseline described in §4.2 ('we instantiate the PointNSP baseline by adopting its official quantization-based VAE reconstruction module coupled with a lightweight global self-attention generator') raises a comparison-fairness concern, but it is not circularity: no fitted parameter of PointRSP is derived from that baseline's evaluation. The limitation statement honestly notes that scalability to ultra-dense or scene-level data is unvalidated, which further supports that the paper does not overclaim a derivation. Overall, no circular step can be identified, so the score is 0.

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

No physically invented entities are introduced; the algorithmic modules (TPA tree, dual-stream generator, GCPE) are not independent postulates with falsifiable handles. The ledger lists the hand-chosen hyperparameters and domain assumptions the method leans on.

free parameters (5)

T, Tl, Tk (number of scales, merge steps, spectral split steps) = Not stated in main text; deferred to supplementary
These control the depth and balance of the non-balanced binary tree and are chosen by hand.
K (number of merged tokens before spectral split) = Not stated
Determines the resolution at which the greedy merge hands off to top-down spectral partitioning.
alpha (geometry-calibrated positional encoding mixing weight) = Not stated; mentioned as a hyperparameter analyzed in supplementary
Directly controls how much early-stage features rely on raw reconstructed coordinates versus structural centers.
Loss weights lambda_L1, lambda_cmf, lambda_vq = Not stated
Weights in Eq. 6 are chosen by hand and influence reconstruction/quantization balance.
Codebook size, latent dimension D, attention layers/heads = Not stated
Core architecture sizes needed to reproduce exact tables are not given.

assumptions (4)

domain assumption The Fiedler-vector sign of the KNN graph Laplacian gives stable and meaningful geometric bisections of every shape in the dataset.
Used to define all split labels via Eq. (4); no experiment measures whether cuts align with semantic part boundaries.
domain assumption A nearest-neighbor graph built on pooled point features is a faithful proxy for intrinsic manifold topology.
This is the 'topology-preserving' premise of Section 3.1; if KNN edges cross thin structures or are noisy, the blueprint does not preserve topology.
domain assumption Feature-space distances from a sparse convolutional encoder are a reliable measure of local geometric similarity for greedy merging.
The bottom-up merge stage selects pairs by 'local feature distances' in Section 3.1; no sensitivity analysis is provided.
domain assumption Residual quantization with codebook commitment converges well enough for the tree structure to remain learnable.
Training relies on VQ losses from [16,25]; the paper does not ablate codebook failure modes.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Learning to Tessellate: Point Cloud Generation via Recursive Spectral Partitioning." pith.science (2026). https://pith.science/paper/BEJ4DSPS

@misc{pith2026260802432,
  author       = {Pith},
  title        = {Pith review of: Learning to Tessellate: Point Cloud Generation via Recursive Spectral Partitioning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BEJ4DSPS}},
  note         = {Machine review of arXiv:2608.02432}
}

read the original abstract

Autoregressive models have emerged as an effective paradigm for point cloud generation. However, most existing approaches rely on heuristic tokenization strategies, such as spatial sorting or stochastic downsampling, which often disrupt intrinsic point cloud topology and weaken the structural coherence of the generated shapes. In this paper, we present PointRSP, an autoregressive framework that reformulates point cloud generation as a topology-preserving tessellation process via recursive spectral partitioning. Instead of constructing token sequences heuristically, we introduce a topology-aware partitioning autoencoder that decomposes an unstructured point cloud into a non-balanced binary tree through a hybrid recursive spectral partitioning strategy. This hierarchical representation provides a deterministic geometric blueprint that preserves topological relationships while capturing multiscale structural dependencies within a quantized latent space. To synthesize shapes in this space, we propose a dual-stream cascaded generator that jointly models structural evolution and feature synthesis. In addition, we design a geometry-calibrated positional encoding mechanism that anchors latent embeddings using multi-scale structural centers, which stabilizes cascaded generation during the early stages of structural formation. Extensive experiments show that PointRSP achieves state-of-the-art performance in generation quality and diversity, demonstrating strong generalization across complex 3D topologies.

Figures

Figures reproduced from arXiv: 2608.02432 by the authors.

**Figure 1.** Comparison between PointRSP and conventional fragmented autoregressive models. Traditional methods (a) generate shapes by merging independently generated subsets, whereas ours (b) follows a topology-preserving tessellation process. Abstract. Autoregressive models have emerged as an effective paradigm for point cloud generation. However, most existing approaches rely on heuristic tokenization strategies, such as spat… view at source ↗

**Figure 2.** Overview of the Topology-Aware Partitioning Autoencoder. The framework extracts a hierarchical feature sequence, applying a topology-preserving tessellated quantization via a recursive spectral partitioning strategy. By dynamically quantizing features within disjoint geometric subsets dictated by a deterministic structural blueprint, the model effectively preserves intrinsic topology during reconstruction. 3 Method… view at source ↗

**Figure 3.** Visualization of the recursive spectral partitioning strategy. The geometric blueprint is constructed through a hybrid approach: a bottom-up greedy merge stage (Tl) captures fine-grained local details, while a top-down spectral split stage (Tk) preserves global topology. Compared to (a) standard KNN-style balanced clustering, (b) our approach explicitly accommodates the non-balanced nature of 3D geometries to maint… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Overview of the cascaded generation from scale t−1 to t. The dual-stream generator employs a structural predictor and a feature predictor, stabilized by a geometrycalibrated positional encoding. The right panel illustrates the dynamic construction of topology-aware p…

**Figure 5.** Figure 5: Qualitative visualization of the stage-wise cascaded generation process across the airplane, car, and chair categories. As the generation progresses from coarse to fine scales, the synthesized points (top row) are progressively anchored and refined by the evolving topo…

**Figure 6.** Figure 6: Qualitative comparison between PointRSP and representative state-of-the-art baselines across single-category generation. Compared to existing autoregressive and diffusion-based methods, PointRSP explicitly preserves intrinsic manifold continuity, resulting in synthesiz…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 3 linked inside Pith

[1]

Achlioptas, P., Diamanti, O., Mitliagkas, I., Guibas, L.: Learning representations and generative models for 3d point clouds. In: Int. Conf. Mach. Learn. pp. 40–49 (2018)

2018
[2]

Bachmann, R., Allardice, J., Mizrahi, D., Fini, E., Kar, O.F., Amirloo, E., El- Nouby, A., Zamir, A., Dehghan, A.: Flextok: Resampling images into 1d token sequences of flexible length. In: Int. Conf. Mach. Learn. pp. 2241–2292 (2025)

2025
[3]

Neural Comput.15(6), 1373–1396 (2003)

Belkin, M., Niyogi, P.: Laplacian eigenmaps for dimensionality reduction and data representation. Neural Comput.15(6), 1373–1396 (2003)

2003
[4]

Chen, G., Wang, M., Yang, Y., Yu, K., Yuan, L., Yue, Y.: Pointgpt: Auto- regressively generative pre-training from point clouds. In: Adv. Neural Inform. Process. Syst. pp. 29667–29679 (2023)

2023
[5]

Chen, M., Radford, A., Child, R., Wu, J., Jun, H., Luan, D., Sutskever, I.: Gener- ative pretraining from pixels. In: Int. Conf. Mach. Learn. pp. 1691–1703 (2020)

2020
[6]

Cheng, A.C., Li, X., Liu, S., Sun, M., Yang, M.H.: Autoregressive 3d shape gener- ation via canonical mapping. In: Eur. Conf. Comput. Vis. pp. 89–104 (2022)

2022
[7]

IEEE Trans

Cheng, Y., Su, J., Jiang, M., Liu, Y.: A novel radar point cloud generation method for robot environment perception. IEEE Trans. Robot.38(6), 3754–3773 (2022)

2022
[8]

In: IEEE Conf

Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12873–12883 (2021)

2021

Show all 42 references

[9]

arXiv preprint arXiv:2505.23660 (2025)

Gao, Z., Shou, M.Z.: D-ar: Diffusion via autoregressive models. arXiv preprint arXiv:2505.23660 (2025)

2025 arXiv
[10]

Girshick, R.: Fast r-cnn. In: Int. Conf. Comput. Vis. pp. 1440–1448 (2015)

2015
[11]

Ho, J., Saharia, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res.23(47), 1–33 (2022)

2022
[12]

In: IEEE Conf

Lee, D., Kim, C., Kim, S., Cho, M., Han, W.S.: Autoregressive image genera- tion using residual quantization. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 11523–11532 (2022)

2022
[13]

IEEE Trans

Liu, B., Yu, Y., Xu, X., Xu, C., Zheng, C., Yang, H., Huang, S., He, S.: Genpoly: Learning generalized and tessellated shape priors via 3d polymorphic evolving. IEEE Trans. Pattern Anal. Mach. Intell. pp. 1–15 (2025)

2025
[14]

Lopez-Paz, D., Oquab, M.: Revisiting classifier two-sample tests. In: Int. Conf. Learn. Represent. pp. 1–12 (2017)

2017
[15]

In: IEEE Conf

Luo, S., Hu, W.: Diffusion probabilistic models for 3d point cloud generation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2837–2845 (2021)

2021
[16]

arXiv preprint arXiv:2409.04410 (2024)

Luo, Z., Shi, F., Ge, Y., Yang, Y., Wang, L., Shan, Y.: Open-magvit2: An open-source project toward democratizing auto-regressive visual generation. arXiv preprint arXiv:2409.04410 (2024)

2024 arXiv
[17]

In: IEEE Conf

Meng, Z., Wang, Q., Dou, Z., Song, Z., Zhou, Z., King, I., Zhao, P.: Pointnsp: Autoregressive 3d point cloud generation with next-scale level-of-detail prediction. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 31450–31461 (2026)

2026
[18]

arXiv preprint arXiv:2511.20041 (2025)

Molodyk, P., Choi, J., Romero, D.W., Liu, M.Y., Chen, Y.: Mfm-point: Multi-scale flow matching for point cloud generation. arXiv preprint arXiv:2511.20041 (2025)

2025
[19]

Razavi, A., Van den Oord, A., Vinyals, O.: Generating diverse high-fidelity images with vq-vae-2. In: Adv. Neural Inform. Process. Syst. pp. 14866–14876 (2019)

2019
[20]

In: IEEE Conf

Ren, Z., Kim, M., Liu, F., Liu, X.: Tiger: Time-varying denoising model for 3d point cloud generation with diffusion process. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 9462–9471 (2024) Point Cloud Generation via Recursive Spectral Partitioning 17

2024
[21]

Shi, F., Luo, Z., Ge, Y., Yang, Y., Shan, Y., Wang, L.: Scalable image tokenization with index backpropagation quantization. In: Int. Conf. Comput. Vis. pp. 16037– 16046 (2025)

2025
[22]

arXiv preprint arXiv:2406.06525 (2024)

Sun, P., Jiang, Y., Chen, S., Zhang, S., Peng, B., Luo, P., Yuan, Z.: Autoregres- sive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525 (2024)

2024 arXiv
[23]

In: IEEE Winter Conf

Sun, Y., Wang, Y., Liu, Z., Siegel, J., Sarma, S.: Pointgrow: Autoregressively learned point cloud generation with self-attention. In: IEEE Winter Conf. Appl. Comput. Vis. pp. 61–70 (2020)

2020
[24]

Tao, Y., Zhang, J., Shi, Y., Zou, D., Zhou, B.: Gsv3d: Gaussian splatting-based geometric distillation with stable video diffusion for single-image 3d object gener- ation. In: Int. Conf. Comput. Vis. pp. 7751–7760 (2025)

2025
[25]

Tian, K., Jiang, Y., Yuan, Z., Peng, B., Wang, L.: Visual autoregressive model- ing: Scalable image generation via next-scale prediction. In: Adv. Neural Inform. Process. Syst. pp. 84839–84865 (2024)

2024
[26]

Syst.35, 10021–10039 (2022)

Vahdat, A., Williams, F., Gojcic, Z., Litany, O., Fidler, S., Kreis, K., et al.: Lion: Latentpointdiffusionmodelsfor3dshapegeneration.Adv.NeuralInform.Process. Syst.35, 10021–10039 (2022)

2022
[27]

Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. In: Adv. Neural Inform. Process. Syst. pp. 6309–6318 (2017)

2017
[28]

Von Luxburg, U.: A tutorial on spectral clustering. Statis. Comput.17(4), 395–416 (2007)

2007
[29]

In: ACM SIGGRAPH

Wang, J., Lin, C., Liu, Y., Xu, R., Dou, Z., Long, X., Guo, H., Komura, T., Wang, W., Li, X.: Pdt: Point distribution transformation with diffusion models. In: ACM SIGGRAPH. pp. 1–11 (2025)

2025
[30]

In: IEEE Conf

Wang, T., Mao, X., Zhu, C., Xu, R., Lyu, R., Li, P., Chen, X., Zhang, W., Chen, K., Xue, T., et al.: Embodiedscan: A holistic multi-modal 3d perception suite towards embodied ai. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19757–19767 (2024)

2024
[31]

Wang, Y., Wang, Z., Wu, Z., Tao, Q., Liao, K., Loy, C.C.: Next visual granularity generation. In: Int. Conf. Learn. Represent. pp. 1–14 (2026)

2026
[32]

ACM Trans

Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. ACM Trans. Graph.38(5), 1–12 (2019)

2019
[33]

Wang, Y., Xian, Z., Chen, F., Wang, T.H., Wang, Y., Fragkiadaki, K., Erickson, Z., Held, D., Gan, C.: Robogen: towards unleashing infinite data for automated robot learning via generative simulation. In: Int. Conf. Mach. Learn. pp. 51936–51983 (2024)

2024
[34]

In: IEEE Conf

Wu, X., Jiang, L., Wang, P.S., Liu, Z., Liu, X., Qiao, Y., Ouyang, W., He, T., Zhao, H.: Point transformer v3: Simpler faster stronger. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4840–4851 (2024)

2024
[35]

Wu, Y., Zheng, W., Zuo, S., Huang, Y., Zhou, J., Lu, J.: Embodiedocc: Embodied 3d occupancy prediction for vision-based online scene understanding. In: Int. Conf. Comput. Vis. pp. 26360–26370 (2025)

2025
[36]

In: IEEE Conf

Yan, X., Lin, L., Mitra, N.J., Lischinski, D., Cohen-Or, D., Huang, H.: Shape- former: Transformer-based shape completion via sparse representation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6239–6249 (2022)

2022
[37]

Sen- sors18(10), 3337 (2018)

Yan, Y., Mao, Y., Li, B.: Second: Sparsely embedded convolutional detection. Sen- sors18(10), 3337 (2018)

2018
[38]

Yang, G., Huang, X., Hao, Z., Liu, M.Y., Belongie, S., Hariharan, B.: Pointflow: 3d point cloud generation with continuous normalizing flows. In: Int. Conf. Comput. Vis. pp. 4541–4550 (2019) 18 M. Sun et al

2019
[39]

arXiv preprint arXiv:2503.05305 (2025)

Yu, H., Luo, H., Yuan, H., Rong, Y., Zhao, F.: Frequency autoregressive image generation with continuous tokens. arXiv preprint arXiv:2503.05305 (2025)

2025
[40]

Zhang, J., Xiong, F., Wang, G., Xu, M.: G3pt: unleash the power of autoregressive modeling in 3d generation via cross-scale querying transformer. In: Int. Jt. Conf. Artif. Intell. pp. 2350–2358 (2025)

2025
[41]

Zheng, C., Lin, Y., Liu, B., Xu, X., Nie, Y., He, S.: Recdreamer: Consistent text- to-3d generation via uniform score distillation. In: Int. Conf. Learn. Represent. pp. 1–14 (2025)

2025
[42]

Zhou, L., Du, Y., Wu, J.: 3d shape generation and completion through point-voxel diffusion. In: Int. Conf. Comput. Vis. pp. 5826–5835 (2021)

2021

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

Achlioptas, P., Diamanti, O., Mitliagkas, I., Guibas, L.: Learning representations and generative models for 3d point clouds. In: Int. Conf. Mach. Learn. pp. 40–49 (2018)

2018

[2] [2]

Bachmann, R., Allardice, J., Mizrahi, D., Fini, E., Kar, O.F., Amirloo, E., El- Nouby, A., Zamir, A., Dehghan, A.: Flextok: Resampling images into 1d token sequences of flexible length. In: Int. Conf. Mach. Learn. pp. 2241–2292 (2025)

2025

[3] [3]

Neural Comput.15(6), 1373–1396 (2003)

Belkin, M., Niyogi, P.: Laplacian eigenmaps for dimensionality reduction and data representation. Neural Comput.15(6), 1373–1396 (2003)

2003

[4] [4]

Chen, G., Wang, M., Yang, Y., Yu, K., Yuan, L., Yue, Y.: Pointgpt: Auto- regressively generative pre-training from point clouds. In: Adv. Neural Inform. Process. Syst. pp. 29667–29679 (2023)

2023

[5] [5]

Chen, M., Radford, A., Child, R., Wu, J., Jun, H., Luan, D., Sutskever, I.: Gener- ative pretraining from pixels. In: Int. Conf. Mach. Learn. pp. 1691–1703 (2020)

2020

[6] [6]

Cheng, A.C., Li, X., Liu, S., Sun, M., Yang, M.H.: Autoregressive 3d shape gener- ation via canonical mapping. In: Eur. Conf. Comput. Vis. pp. 89–104 (2022)

2022

[7] [7]

IEEE Trans

Cheng, Y., Su, J., Jiang, M., Liu, Y.: A novel radar point cloud generation method for robot environment perception. IEEE Trans. Robot.38(6), 3754–3773 (2022)

2022

[8] [8]

In: IEEE Conf

Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 12873–12883 (2021)

2021