pith. sign in

arxiv: 2605.29655 · v2 · pith:LSYBBQYBnew · submitted 2026-05-28 · 💻 cs.CV · cs.GR

SuperVoxelGPT: Adaptive and Ordered 3D Tokenization for Autoregressive Shape Generation

Yuan Li , Congyi Zhang , Xifeng Gao , Xiaohu Guo This is my paper

Pith reviewed 2026-06-29 08:49 UTC · model grok-4.3

classification 💻 cs.CV cs.GR
keywords 3D shape generationautoregressive modelingsupervoxel tokenizationmultimodal large language modelsVoronoi tessellationadaptive partitioningtext-to-3Dshape representation
0
0 comments X

The pith

Adaptive supervoxel tokenization cuts 3D sequence lengths to 12.8% of uniform voxels for autoregressive generation.

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

The paper seeks to resolve the incompatibility between compact but unordered set-based 3D representations and ordered but lengthy grid-based ones for use in autoregressive multimodal models. It predicts a coarse saliency map from the input prompt and applies saliency-guided centroidal Voronoi tessellation to build an adaptive partition of supervoxels, with finer cells in complex regions and larger cells in smooth areas. The resulting sequence is both short and deterministically ordered, allowing a SuperVoxelVAE encoder and a fine-tuned MLLM to perform autoregressive token prediction. On the Trellis-500K benchmark this produces state-of-the-art shapes at much lower computational cost.

Core claim

SuperVoxelGPT resolves the structural trade-off between set-based and grid-based 3D tokenizations through adaptive and deterministically ordered supervoxel tokenization. Given a prompt, a coarse geometric saliency distribution is predicted and used to drive saliency-guided centroidal Voronoi tessellation that allocates fine-grained cells to complex regions and larger cells to smooth regions. The resulting compact, ordered supervoxel layout is encoded by a SuperVoxelVAE and generated autoregressively by a fine-tuned pretrained MLLM.

What carries the argument

saliency-guided centroidal Voronoi tessellation that produces shape-adaptive supervoxel partitions with deterministic ordering

If this is right

  • Token sequence length is reduced to 12.8% of uniform voxel tokenization
  • State-of-the-art generation quality is achieved on Trellis-500K
  • An average 10x speedup is obtained over prior methods
  • Autoregressive prediction proceeds stably without ordering ambiguities

Where Pith is reading between the lines

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

  • The same adaptive partitioning could be applied to point-cloud or mesh inputs to test whether sequence-length reductions transfer to other 3D modalities.
  • Replacing the separate saliency-prediction stage with a fully end-to-end learned module might further reduce preprocessing overhead.
  • The fixed ordering property could support direct concatenation of supervoxel sequences with 2D image token streams for joint multimodal training.

Load-bearing premise

That a coarse geometric saliency distribution predicted from the prompt can be used to drive centroidal Voronoi tessellation into a shape-adaptive supervoxel partition whose resulting token sequence is both compact and deterministically ordered enough for stable autoregressive prediction without introducing new ambiguities.

What would settle it

If autoregressive models trained on the resulting supervoxel sequences exhibit higher inconsistency or lower shape quality than equivalent models trained on uniform voxel sequences, the claim that the adaptive partitions preserve sufficient structure would be refuted.

Figures

Figures reproduced from arXiv: 2605.29655 by Congyi Zhang, Xiaohu Guo, Xifeng Gao, Yuan Li.

Figure 1
Figure 1. Figure 1: We propose a new two-stage MLLM framework for high-resolution 3D genera￾tion. The model first predicts a saliency map (a), which is then transformed into our SuperVoxel structure via a customized 3D CVT (b), enabling the extraction of com￾pact and ordered tokens. Finally, our autoregressive model, SuperVoxelGPT, generates high-resolution 3D geometry based on this representation (c). Abstract. Autoregressiv… view at source ↗
Figure 2
Figure 2. Figure 2: (a) Original mesh, (b) Uniform voxel structure, (c) Mesh saliency, (d) Volume saliency (e) Supervoxel structure 1 Introduction In recent years, Multimodal Large Language Models (MLLMs) [44] have achieved remarkable success across a wide range of generative tasks [1]. Built upon pre￾trained large language models and fine-tuned on multimodal data, modern MLLMs can jointly reason over text and images, advanci… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of SuperVoxelGPT. SuperVoxelGPT employs two stages for 3D asset generation: (a) Prompt-to-supervoxel structure. In the Prompt-to-supervoxel stage, we first predict the coarse saliency distribution of a shape via a lightweight MaskGIT model, then use Saliency-guided Centroidal Voronoi Tessellation to partition the space into adaptive supervoxels based on the saliency distribution. (b) Supervoxel-to… view at source ↗
Figure 4
Figure 4. Figure 4: Overview of two VAE structures. (a) Saliency VQ-VAE, (b) SuperVoxelVAE. and supervoxel positions, then decode the tokens into the final 3D shape. We describe the Prompt-to-supervoxel structure stage in Sec. 3.1 and the supervoxel￾to-shape generation stage in Sec. 3.2. 3.1 Prompt-to-Supervoxel Structure Stage This stage predicts where to allocate tokens based on the geometric complex￾ity distribution of the… view at source ↗
Figure 6
Figure 6. Figure 6: Effect of saliency-driven CVT. The supervoxel partition adapts to geometric complexity, with denser cells near detailed features and sparser cells in smooth regions. TRELLIS2 [41] and add an additional L1 loss to supervise the saliency values. In this way, a saliency volume is mapped into 1024 tokens. Saliency Volume MaskGIT. To avoid extensive autoregressive and de￾noising steps, we adopt the MaskGIT arch… view at source ↗
Figure 5
Figure 5. Figure 5: Compression ratio c as a function of max supervoxel size K for different saliency thresholds t. The monotonic decrease en￾ables efficient parameter selection for target compression ratios. Our goal is to define a piecewise￾linear target size field from saliency values to final supervoxel sizes: in geometrically complex regions (high saliency), we maintain unit-size su￾pervoxels with a 1:1 correspondence to… view at source ↗
Figure 7
Figure 7. Figure 7: Architecture of the MLLM generation module. The MLLM autoregressively generates the token sequence conditioned on the multimodal prompt, which is then decoded into the final 3D shape. 3.2 Supervoxel-to-Shape Generation Given a text or image prompt and the supervoxel structure from the previous stage, this stage generates the corresponding 3D shapes. It consists of two com￾ponents: (1) SuperVoxelVAE, which … view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison on image-to-3D generation. Given an input image, we compare the generated 3D shapes from CraftsMans, TRELLIS, Direct3D-S2, TREL￾LIS2, and our method (SV, i.e., supervoxel structure). would reduce the compression ratio, as shown in [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
read the original abstract

Autoregressive multimodal large language models (MLLMs) enable 3D generation but struggle to scale to high-resolution shapes due to inadequate 3D tokenizations. Compact set-based representations discard deterministic spatial ordering, leading to ambiguous sequence prediction, while uniform or octree-based voxel grids preserve ordering at the cost of severe redundancy and excessively long sequences. This structural trade-off limits stable and efficient autoregressive 3D generation. We present SuperVoxelGPT, a representation-first framework that resolves this tension through adaptive and deterministically ordered supervoxel tokenization. Given a prompt, we first predict a coarse geometric saliency distribution and construct a shape-adaptive supervoxel partition using saliency-guided centroidal Voronoi tessellation, allocating fine-grained cells to complex regions and larger cells to smooth regions. Conditioned on the text and ordered supervoxel layout, we introduce a SuperVoxelVAE and fine-tune a pretrained MLLM to autoregressively generate supervoxel tokens. Experiments on Trellis-500K show that SuperVoxelGPT reduces token sequence length to 12.8% of uniform voxel tokenization while achieving state-of-the-art generation quality and an average 10$\times$ speedup over prior methods.

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

2 major / 1 minor

Summary. The paper introduces SuperVoxelGPT, a framework for 3D shape generation that addresses limitations in existing tokenizations for autoregressive MLLMs. It predicts a coarse geometric saliency from the prompt, uses saliency-guided centroidal Voronoi tessellation to create adaptive supervoxels (finer in complex regions), imposes a deterministic order on the cells, and then uses a SuperVoxelVAE and fine-tuned MLLM to generate the token sequence autoregressively. On the Trellis-500K dataset, it claims a reduction in token sequence length to 12.8% of uniform voxel tokenization, state-of-the-art generation quality, and an average 10x speedup.

Significance. If the results are substantiated, this approach could significantly improve the scalability of autoregressive 3D generation by providing a compact yet ordered representation that avoids the ambiguities of set-based methods and the redundancy of grid-based ones. The adaptive allocation of resolution based on saliency is a promising direction for efficient high-resolution modeling.

major comments (2)
  1. [Abstract and Experiments] The quantitative claims (12.8% token length, SOTA quality, 10x speedup on Trellis-500K) are presented without any supporting tables, baseline comparisons, ablation studies, dataset statistics, or error bars, making it impossible to evaluate whether the data support the central claims.
  2. [Method (saliency-guided CVT and ordering)] The deterministic ordering of supervoxel tokens is derived from a learned saliency prediction followed by CVT; the manuscript provides no analysis, perturbation tests, or ablations demonstrating that the resulting sequence order is stable under small variations in the predicted saliency map. This stability is load-bearing for the claim that the representation avoids introducing new ambiguities for autoregressive prediction.
minor comments (1)
  1. [Abstract] The abstract refers to 'Trellis-500K' without a citation or brief description of the dataset.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback. We address each major comment below and describe the revisions planned for the manuscript.

read point-by-point responses

  1. Referee: [Abstract and Experiments] The quantitative claims (12.8% token length, SOTA quality, 10x speedup on Trellis-500K) are presented without any supporting tables, baseline comparisons, ablation studies, dataset statistics, or error bars, making it impossible to evaluate whether the data support the central claims.

    Authors: We agree that the abstract and current Experiments section do not include the detailed supporting evidence needed to substantiate the claims. In the revised manuscript we will expand the Experiments section to include full tables with baseline comparisons, ablation studies on key components, dataset statistics for Trellis-500K, and error bars computed over multiple runs, thereby providing the necessary quantitative support for the reported token-length reduction, generation quality, and speedup. revision: yes

  2. Referee: [Method (saliency-guided CVT and ordering)] The deterministic ordering of supervoxel tokens is derived from a learned saliency prediction followed by CVT; the manuscript provides no analysis, perturbation tests, or ablations demonstrating that the resulting sequence order is stable under small variations in the predicted saliency map. This stability is load-bearing for the claim that the representation avoids introducing new ambiguities for autoregressive prediction.

    Authors: We concur that empirical verification of ordering stability under saliency variations is important to support the autoregressive modeling claim. In the revision we will add a dedicated analysis subsection containing perturbation tests: small controlled variations will be introduced to the predicted saliency maps, after which we will quantify changes in the resulting CVT partitions and token sequences using appropriate stability metrics, together with qualitative examples. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained.

full rationale

The paper presents a new representation (saliency-guided CVT supervoxel partition + SuperVoxelVAE + MLLM fine-tuning) whose claimed benefits (shorter sequences, stable AR ordering, SOTA quality) are not shown to reduce by construction to fitted inputs or prior self-citations. The abstract and method description treat the tokenization as an independent design choice rather than a re-expression of existing quantities. No equations or steps equate the output ordering or performance to the input saliency prediction by definition. This is the expected non-finding for a representation-first paper.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Only the abstract is available, so the ledger is necessarily incomplete. The central claim rests on the unverified assumption that saliency-guided CVT yields a usable ordered partition and that the reported speed/quality numbers are reproducible.

free parameters (1)
  • saliency prediction network parameters
    The coarse geometric saliency distribution is predicted by a learned model whose weights are fitted during training.
axioms (1)
  • domain assumption Centroidal Voronoi tessellation on a saliency field produces a deterministically ordered supervoxel layout suitable for autoregressive modeling
    Invoked when the paper states that the partition preserves deterministic spatial ordering.

pith-pipeline@v0.9.1-grok · 5754 in / 1363 out tokens · 32847 ms · 2026-06-29T08:49:53.804473+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

49 extracted references · 21 canonical work pages · 5 internal anchors

  1. [1]

    arXiv preprint arXiv:2402.12451 (2024)

    Caffagni, D., Cocchi, F., Barsellotti, L., Moratelli, N., Sarto, S., Baraldi, L., Cornia, M., Cucchiara, R.: The revolution of multimodal large language models: a survey. arXiv preprint arXiv:2402.12451 (2024)

    work page arXiv 2024

  2. [2]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked genera- tive image transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11315–11325 (2022)

    2022

  3. [3]

    arXiv preprint arXiv:2503.20519 (2025)

    Chen, J., Zhu, L., Hu, Z., Qian, S., Chen, Y., Wang, X., Lee, G.H.: Mar-3d: Pro- gressive masked auto-regressor for high-resolution 3d generation. arXiv preprint arXiv:2503.20519 (2025)

    work page arXiv 2025

  4. [4]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Chen, R., Zhang, J., Liang, Y., Luo, G., Li, W., Liu, J., Li, X., Long, X., Feng, J., Tan, P.: Dora: Sampling and benchmarking for 3d shape variational auto-encoders. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 16251–16261 (2025)

    2025

  5. [5]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Chen, Y., Wang, Y., Luo, Y., Wang, Z., Chen, Z., Zhu, J., Zhang, C., Lin, G.: Meshanything v2: Artist-created mesh generation with adjacent mesh tokenization. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 13922–13931 (2025)

    2025

  6. [6]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Chen, Y., Lan, Y., Zhou, S., Wang, T., Pan, X.: Sar3d: Autoregressive 3d object generation and understanding via multi-scale 3d vqvae. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 28371–28382 (2025)

    2025

  7. [7]

    In: ICLR (2024)

    Delétang, G., Ruoss, A., Duquenne, P.A., Catt, E., Genewein, T., Mattern, C., Grau-Moya, J., Wenliang, L.K., Aitchison, M., Orseau, L., Hutter, M., Veness, J.: Language modeling is compression. In: ICLR (2024)

    2024

  8. [8]

    In: Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision

    Deng, K., Liu, H.T.D., Zhu, Y., Sun, X., Shang, C., Bhat, K.S., Ramanan, D., Zhu, J.Y., Agrawala, M., Zhou, T.: Efficient autoregressive shape generation via octree-based adaptive tokenization. In: Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision. pp. 11685–11696 (2025)

    2025

  9. [9]

    SIAM review41(4), 637–676 (1999)

    Du, Q., Faber, V., Gunzburger, M.: Centroidal voronoi tessellations: Applications and algorithms. SIAM review41(4), 637–676 (1999)

    1999

  10. [10]

    Numerical Mathematics: Theory, Methods and Applications 3(2), 119–142 (2010)

    Du, Q., Gunzburger, M., Ju, L.: Advances in studies and applications of centroidal voronoi tessellations. Numerical Mathematics: Theory, Methods and Applications 3(2), 119–142 (2010)

    2010

  11. [11]

    Fan, H., Su, H., Guibas, L.J.: A point set generation network for 3d object recon- structionfromasingleimage.In:ProceedingsoftheIEEEConferenceonComputer Vision and Pattern Recognition. pp. 605–613 (2017)

    2017

  12. [12]

    arXiv preprint arXiv:2503.21732 (2025)

    He, X., Zou, Z.X., Chen, C.H., Guo, Y.C., Liang, D., Yuan, C., Ouyang, W., Cao, Y.P., Li, Y.: Sparseflex: High-resolution and arbitrary-topology 3d shape modeling. arXiv preprint arXiv:2503.21732 (2025)

    work page arXiv 2025

  13. [13]

    arXiv preprint arXiv:2511.00763 (2025)

    Hou, W., Zhou, L., Hu, H.Y., Chen, Y., You, Y.Z., Qi, X.L.: How focused are llms? a quantitative study via repetitive deterministic prediction tasks. arXiv preprint arXiv:2511.00763 (2025)

    work page arXiv 2025

  14. [14]

    Qwen2.5-Coder Technical Report

    Hui, B., Yang, J., Cui, Z., Yang, J., Liu, D., Zhang, L., Liu, T., Zhang, J., Yu, B., Lu, K., et al.: Qwen2. 5-coder technical report. arXiv preprint arXiv:2409.12186 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  15. [15]

    arxiv preprint arXiv:2512.21185 (2025) 16 Y

    Jia, T., Yan, D., Hao, D., Li, Y., Zhang, K., He, X., Li, L., Chen, J., Jiang, L., Yin, Q., Quan, L., Chen, Y.C., Yuan, L.: Ultrashape 1.0: High-fidelity 3d shape gener- ation via scalable geometric refinement. arxiv preprint arXiv:2512.21185 (2025) 16 Y. Li et al

    work page arXiv 2025

  16. [16]

    low-resource

    Jiang, Z., Yang, M., Tsirlin, M., Tang, R., Dai, Y., Lin, J.: “low-resource” text classification: A parameter-free classification method with compressors. In: Find- ings of the Association for Computational Linguistics: ACL 2023. pp. 6810–6828. Association for Computational Linguistics (2023)

    2023

  17. [17]

    Hunyuan3D 2.5: Towards High-Fidelity 3D Assets Generation with Ultimate Details

    Lai, Z., Zhao, Y., Liu, H., Zhao, Z., Lin, Q., Shi, H., Yang, X., Yang, M., Yang, S., Feng, Y., et al.: Hunyuan3d 2.5: Towards high-fidelity 3d assets generation with ultimate details. arXiv preprint arXiv:2506.16504 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  18. [18]

    In: ACM SIGGRAPH 2005 Papers, pp

    Lee, C.H., Varshney, A., Jacobs, D.W.: Mesh saliency. In: ACM SIGGRAPH 2005 Papers, pp. 659–666 (2005)

    2005

  19. [19]

    ACM Transactions on Graphics (TOG)29(4), 1–11 (2010)

    Lévy, B., Liu, Y.: L p centroidal voronoi tessellation and its applications. ACM Transactions on Graphics (TOG)29(4), 1–11 (2010)

    2010

  20. [20]

    arXiv preprint arXiv:2405.14979 (2024)

    Li, W., Peng, J., Chen, H., Gu, L., Wang, Q.: Craftsman: High-fidelity mesh gen- eration with 3d native generation and interactive geometry refiner. arXiv preprint arXiv:2405.14979 (2024)

    work page arXiv 2024

  21. [21]

    arXiv preprint arXiv:2505.14521 (2025)

    Li, Z., Wang, Y., Zheng, H., Luo, Y., Wen, B.: Sparc3d: Sparse representa- tion and construction for high-resolution 3d shapes modeling. arXiv preprint arXiv:2505.14521 (2025)

    work page arXiv 2025

  22. [22]

    arXiv preprint arXiv:2505.19901 (2025)

    Liu, P., Ren, X., Liu, F., Xie, Q., Zheng, Q., Zhang, Y., Lu, H., Yang, Y.: Dynamic-i2v: Exploring image-to-video generation models via multimodal llm. arXiv preprint arXiv:2505.19901 (2025)

    work page arXiv 2025

  23. [23]

    arXiv preprint arXiv:2309.13638 (2023)

    McCoy, R.T., Yao, S., Friedman, D., Hardy, M., Griffiths, T.L.: Embers of au- toregression: Understanding large language models through the problem they are trained to solve. arXiv preprint arXiv:2309.13638 (2023)

    work page arXiv 2023

  24. [24]

    Finite Scalar Quantization: VQ-VAE Made Simple

    Mentzer, F., Minnen, D., Agustsson, E., Tschannen, M.: Finite scalar quantization: Vq-vae made simple. arXiv preprint arXiv:2309.15505 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  25. [25]

    Visualization and Mathematics III pp

    Meyer, M., Desbrun, M., Schröder, P., Barr, A.H.: Discrete differential-geometry operators for triangulated 2-manifolds. Visualization and Mathematics III pp. 35– 57 (2003)

    2003

  26. [26]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Pun, A., Deng, K., Liu, R., Ramanan, D., Liu, C., Zhu, J.Y.: Generating physically stable and buildable brick structures from text. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14798–14809 (2025)

    2025

  27. [27]

    In: SIGGRAPH Asia 2020 Courses (2020)

    Ravi,N.,Reizenstein,J.,Novotny,D.,Gordon,T.,Lo,W.Y.,Johnson,J.,Gkioxari, G.: Accelerating 3d deep learning with pytorch3d. In: SIGGRAPH Asia 2020 Courses (2020)

    2020

  28. [28]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Ren, X., Huang, J., Zeng, X., Museth, K., Fidler, S., Williams, F.: Xcube: Large- scale 3d generative modeling using sparse voxel hierarchies. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4209–4219 (2024)

    2024

  29. [29]

    In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

    Santilli, A., Severino, S., Postolache, E., Maiorca, V., Mancusi, M., Marin, R., Rodolà, E.: Accelerating transformer inference for translation via parallel decoding. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 12336–12355 (2023)

    2023

  30. [30]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Siddiqui, Y., Alliegro, A., Artemov, A., Tommasi, T., Sirigatti, D., Rosov, V., Dai, A., Nießner, M.: Meshgpt: Generating triangle meshes with decoder-only trans- formers. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 19615–19625 (2024)

    2024

  31. [31]

    arXiv preprint arXiv:2507.02477 (2025)

    Song, G., Zhao, Z., Weng, H., Zeng, J., Jia, R., Gao, S.: Mesh silksong: Auto- regressive mesh generation as weaving silk. arXiv preprint arXiv:2507.02477 (2025)

    work page arXiv 2025

  32. [32]

    ACM Transactions On Graphics (TOG)33(1), 1–17 (2014) SuperVoxelGPT 17

    Song, R., Liu, Y., Martin, R.R., Rosin, P.L.: Mesh saliency via spectral processing. ACM Transactions On Graphics (TOG)33(1), 1–17 (2014) SuperVoxelGPT 17

    2014

  33. [33]

    arXiv preprint arXiv:2409.18114 (2024)

    Tang, J., Li, Z., Hao, Z., Liu, X., Zeng, G., Liu, M.Y., Zhang, Q.: Edgerun- ner: Auto-regressive auto-encoder for artistic mesh generation. arXiv preprint arXiv:2409.18114 (2024)

    work page arXiv 2024

  34. [34]

    Tian, K., Jiang, Y., Yuan, Z., Peng, B., Wang, L.: Visual autoregressive modeling: Scalableimagegenerationvianext-scaleprediction.Advancesinneuralinformation processing systems37, 84839–84865 (2024)

    2024

  35. [35]

    Advances in neural information pro- cessing systems30(2017)

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. Advances in neural information pro- cessing systems30(2017)

    2017

  36. [36]

    In: Proceedings of the 32nd ACM International Conference on Multimedia

    Wang, Z., Wang, L., Zhao, Z., Wu, M., Lyu, C., Li, H., Cai, D., Zhou, L., Shi, S., Tu, Z.: Gpt4video: A unified multimodal large language model for lnstruction- followed understanding and safety-aware generation. In: Proceedings of the 32nd ACM International Conference on Multimedia. pp. 3907–3916 (2024)

    2024

  37. [37]

    arXiv preprint arXiv:2406.12998 (2024)

    Wang, Z., Guo, J., Chen, Z., Zhu, J., Zhang, C.: Llama-mesh: Unifying 3d mesh generation with language models. arXiv preprint arXiv:2406.12998 (2024)

    work page arXiv 2024

  38. [38]

    In: Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Confer- ence Conference Papers

    Wei, S.T., Wang, R.H., Zhou, C.Z., Chen, B., Wang, P.S.: Octgpt: Octree-based multiscale autoregressive models for 3d shape generation. In: Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Confer- ence Conference Papers. pp. 1–11 (2025)

    2025

  39. [39]

    arXiv preprint arXiv:2505.17412 (2025)

    Wu, S., Lin, Y., Zhang, F., Zeng, Y., Yang, Y., Bao, Y., Qian, J., Zhu, S., Cao, X., Torr, P., et al.: Direct3d-s2: Gigascale 3d generation made easy with spatial sparse attention. arXiv preprint arXiv:2505.17412 (2025)

    work page arXiv 2025

  40. [40]

    Advances in Neural Information Processing Systems35, 33330–33342 (2022)

    Wu, X., Lao, Y., Jiang, L., Liu, X., Zhao, H.: Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems35, 33330–33342 (2022)

    2022

  41. [41]

    Native and Compact Structured Latents for 3D Generation

    Xiang, J., Chen, X., Xu, S., Wang, R., Lv, Z., Deng, Y., Zhu, H., Dong, Y., Zhao, H., Yuan, N.J., et al.: Native and compact structured latents for 3d generation. arXiv preprint arXiv:2512.14692 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  42. [42]

    Structured 3D Latents for Scalable and Versatile 3D Generation

    Xiang, J., Lv, Z., Xu, S., Deng, Y., Wang, R., Zhang, B., Chen, D., Tong, X., Yang, J.: Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  43. [43]

    arXiv preprint arXiv:2305.08275 (2024)

    Xue, L., Yu, N., Zhang, S., Panagopoulou, A., Li, J., Martín-Martín, R., Wu, J., Xiong, C., Xu, R., Niebles, J.C., et al.: Ulip-2: Towards scalable multimodal pre- training for 3d understanding. arXiv preprint arXiv:2305.08275 (2024)

    work page arXiv 2024

  44. [44]

    National Science Review11(12), nwae403 (2024)

    Yin, S., Fu, C., Zhao, S., Li, K., Sun, X., Xu, T., Chen, E.: A survey on multimodal large language models. National Science Review11(12), nwae403 (2024)

    2024

  45. [45]

    ACM Trans

    Zhang, B., Tang, J., Nießner, M., Wonka, P.: 3dshape2vecset: A 3d shape repre- sentation for neural fields and generative diffusion models. ACM Trans. Graph. 42(4) (jul 2023).https://doi.org/10.1145/3592442,https://doi.org/10. 1145/3592442

    work page doi:10.1145/3592442 2023

  46. [46]

    arXiv preprint arXiv:2505.18947 (2025)

    Zhang, Z., Shi, Y., Yang, L., Ni, S., Ye, Q., Wang, J.: Openhoi: Open-world hand- object interaction synthesis with multimodal large language model. arXiv preprint arXiv:2505.18947 (2025)

    work page arXiv 2025

  47. [47]

    In: Symposium on interactive 3D graphics and games

    Zheng, J., Tan, T.S.: Computing centroidal voronoi tessellation using the gpu. In: Symposium on interactive 3D graphics and games. pp. 1–9 (2020)

    2020

  48. [48]

    In: European conference on computer vision

    Zhou, Q.Y., Park, J., Koltun, V.: Fast global registration. In: European conference on computer vision. pp. 766–782. Springer (2016)

    2016

  49. [49]

    IEEE Transactions on Information Theory23(3), 337–343 (1977) 18 Y

    Ziv, J., Lempel, A.: A universal algorithm for sequential data compression. IEEE Transactions on Information Theory23(3), 337–343 (1977) 18 Y. Li et al. SuperVoxelGPT: Supplementary Material A Metrics Calculation We provide detailed definitions of the evaluation metrics used in the main paper. We first describe the shape alignment procedure (Sec. A.1), wh...

    1977