Pith. sign in

REVIEW 4 major objections 4 minor 5 cited by

Efficient Part-level 3D Object Generation via Dual Volume Packing

T0 review · 4 major / 4 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 The paper claims that part-level 3D generation can be done end-to-end from a single image by packing all parts into two volumes, yielding complete separable parts in about 30 seconds regardless of part count.

desk verdict Novel dual-volume packing idea, but the 'arbitrary complete parts' claim outruns the evidence; deserves review with requests for contraction statistics and part-level metrics. read the letter →

arxiv 2506.09980 v1 pith:BOH7L5FQ submitted 2025-06-11 cs.CV

Jiaxiang Tang , Ruijie Lu , Zhaoshuo Li , Zekun Hao , Xuan Li , Fangyin Wei , Shuran Song , Gang Zeng
show 2 more authors
Ming-Yu Liu Tsung-Yi Lin
This is my paper · ORCID
classification cs.CV
keywords part-level3Dgenerationdualvolumepackingbipartitegraphcontractionimage-to-3Dlatentdiffusionmeshdecompositionsingle-viewreconstructionrectifiedflow
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

Most image-to-3D generators output one fused mesh, so individual parts cannot be edited or animated. This paper argues that a single image can instead be mapped directly to an object made of complete, semantically meaningful parts, with no limit on how many parts there are. The key move is to organize the parts into two complementary volumes so that parts that touch are never in the same volume; within a volume the parts are disjoint and therefore separate cleanly during mesh extraction. If the claim holds, part-level generation stops being a multi-stage pipeline with segmentation priors and per-part completion, and becomes a fixed-cost one-step generation task.

What carries the argument

Dual volume packing: the part-connectivity graph $G=(V,E)$ is reduced by greedy edge contraction (contract, inside each odd cycle, the edge with the greatest penetration depth, then repeat) until the graph is bipartite, i.e. $\chi(G)=2$. The two color classes are packed into two signed-distance volumes; because no two contacting parts share a volume, iso-surface extraction separates them. The two volumes are encoded into two latent codes that a rectified flow model denoises jointly, so the output length is fixed at two volumes while the number of parts stays variable.

What would settle it

Take an object whose part-connectivity graph has an odd cycle (for example three sub-parts that all touch each other), generate it from an image with this method, and check whether all three parts appear as separate meshes; if any pair comes out fused, the two-volume packing is the reason.

Watch

Extended reading notes

Core claim

The paper's central claim is that a single-view image can be fed end-to-end into a latent denoising model whose output is a pair of packed volumes, and that this pair encodes enough information to reconstruct the full object as an arbitrary number of complete, interleaved mesh parts. The authors build a part-connectivity graph in which vertices are parts and edges indicate contact, then contract edges heuristically to make the graph bipartite, so two colors suffice and every edge joins parts that live in different volumes. The VAE is trained on the two packed volumes, and a flow model denoises the two latent codes jointly, conditioned on image features, with a learned embedding on the second code to stop duplicated parts. At inference both codes are decoded, each volume yields disjoint mesh parts, and the parts assemble into the whole object in about 30 seconds regardless of how many parts the object has.

Load-bearing premise

Most real objects' part-contact patterns can be rearranged into two groups so that no two touching parts land in the same group; when three parts all touch one another, two volumes cannot keep them apart and some parts will be fused or lost.

Editorial extensions

If this is right

  • Inference time stays effectively constant as the number of parts grows, because generation always processes two latent codes rather than one code per part.
  • Generated objects come with complete, separable parts, so articulation, replacement, and other part-level edits become possible without a separate segmentation step.
  • Removing the reliance on 2D segmentation priors eliminates a source of error propagation and a multi-minute preprocessing stage in existing part-level pipelines.
  • The dual-latent formulation is compatible with existing latent denoising models, suggesting the packing idea transfers to other base architectures.
  • The model retains competitive image-to-3D quality while adding structure, trading a small amount of fidelity for decomposition that fused-shape generators lack.

Reading between the lines

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

  • If the bipartite assumption holds on a large fraction of real objects, the method may scale to objects with hundreds of parts because the two volumes impose no cap on part count.
  • The same packing idea should extend to more than two volumes; for graphs whose connectivity is planar, four volumes would suffice by the four-color theorem, which the paper itself notes as future work.
  • Components of the pipeline (graph contraction, volume packing, joint denoising) could be reused for other structured outputs such as layered, multi-material, or interlocking geometry.
  • A user-controllable version might condition generation on a requested part granularity or segmentation mask, turning the current unpredictable granularity into a dial.
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

4 major / 4 minor

Summary. The paper presents a single-image-to-3D generation framework that outputs separable semantic parts. It constructs a part-contact graph from GLB scene graphs, greedily contracts edges to make the graph bipartite, packs the resulting two color classes into two SDF volumes, and trains a VecSet-based VAE plus a rectified flow model to generate the two latent codes jointly from a DINOv2 image embedding. At inference, the two decoded volumes contain non-contacting parts that are assembled into the full object. The paper reports qualitative comparisons with TripoSG, Hi3DGen, and HoloPart, a fused-shape quantitative comparison on 40 images (Table 1), ablation figures, and a limitations section.

Significance. If the method's claims are substantiated, the contribution is practically significant: it offers constant inference time independent of part count, no dependence on 2D/3D segmentation priors at inference, and a fixed-length latent representation for variable-part objects. The technical core, dual volume packing via bipartite contraction, is simple and builds on strong existing components (CLAY, Dora, Trellis, Hunyuan3D-2). However, the current evidence is insufficient: the only quantitative results are whole-mesh similarity scores that cannot validate part-level claims, and the bipartite contraction step is neither analyzed nor benchmarked. The paper is honest about its limitations, but those limitations directly affect the central claim.

major comments (4)
  1. [Sec 3.1, Algorithm 1] The contraction step is load-bearing and is not evaluated. To obtain a bipartite graph, Algorithm 1 merges the parts connected by the maximum-penetration edge in each odd cycle. Merging two parts destroys the boundary between them, so the output cannot contain all original parts whenever the graph is non-bipartite. The paper reports no statistics on how many training meshes require contraction, how many edges are contracted per mesh, how many ground-truth parts are lost, or how often the heuristic produces semantically meaningless unions. Figure 8 gives only the part-count distribution, not contraction statistics. Without these numbers, the abstract's claim of 'an arbitrary number of complete and semantically meaningful parts' is not supported. Please add contraction statistics and a quantitative part-retention evaluation on a labeled subset.
  2. [Sec 4.5, Fig 10] The admitted limitation that 'three mutually contacting parts cannot be represented using only two volumes' is not a rare edge case; it is a structural restriction of the representation. For common objects such as vehicles or furniture, part-contact graphs frequently contain triangles (e.g., wheel-body-axle contact), so at least one semantic boundary must be contracted. Figure 10 directly shows this failure mode: wheels merged with the car body and missing internal structures. The manuscript should quantify the fraction of objects affected, either in the dataset or in generated outputs, and either restrict the claims to bipartite-approximable graphs or add a mechanism (e.g., more volumes) for the remaining cases.
  3. [Table 1, Sec 4.3] The only quantitative evaluation is not diagnostic for part-level generation. Table 1 reports ULIP/ULIP-2/Uni3D whole-shape cosine similarity on 40 images with no error bars or significance testing. The part-level fine-tuned model scores below the fused pretrained model on all three metrics, yet the text describes this as a 'favorable trade-off'. These metrics cannot detect missing parts, fused parts, or incorrect part counts; a single fused blob can score well. The paper needs part-level quantitative metrics (part-count accuracy, part IoU or mIoU, part completeness, contact-graph correctness) and a quantitative comparison against HoloPart/PartField, rather than the manual selection described in Sec 4.2. This is necessary to support the claimed quality, diversity, and generalization advantages.
  4. [Sec 3.1, Algorithm 1] The cycle-elimination algorithm is not sound as stated. The text says 'a depth-first search (DFS) algorithm to identify all cycles', but a standard DFS does not enumerate all simple cycles; it produces a cycle basis. If the algorithm misses odd cycles, the termination condition 'repeat until no odd cycles remain' can be false. Please either use a correct simple-cycle enumeration method or clearly state that the cycle set is an approximation and validate the contraction result empirically (e.g., by verifying bipartiteness of the final graph).
minor comments (4)
  1. [Sec 3.2] The fallback to connected components as part annotations when a mesh has a single geometry node can produce parts that are arbitrary fragments caused by UV seams; the empirical post-processing rules are not validated. Please report the fraction of meshes for which the fallback is used and at least qualitatively assess annotation quality.
  2. [Figure 8] The axes of Figure 8 are not labeled; it is unclear whether the curve is a histogram or cumulative distribution, and the '≥' label is ambiguous.
  3. [Eq. (1)] The data-filtering criterion in Eq. (1) uses o1 and o2 without defining them in the main text beyond 'occupancy ratios'; please define precisely (e.g., fraction of voxels with negative SDF).
  4. [Sec 4.4] The ablation study in Figure 7 is qualitative only; adding quantitative scores (e.g., part-count accuracy or shape metrics) would make the claims about the part embedding and data filtering verifiable.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the dual-volume packing and generation pipeline is a trained system evaluated on held-out images, with no fitted parameter or self-citation chain that reduces the central claim to its inputs.

full rationale

The paper's derivation chain is a standard image-conditioned latent-diffusion pipeline: it packs scene-graph parts into two volumes via a heuristic bipartite-contraction preprocessing step, trains a VAE and rectified-flow model on those packed volumes, and evaluates the resulting generator on held-out images with ULIP and Uni3D metrics. Nothing in this chain is defined in terms of the final output: the bipartite contraction uses only training mesh connectivity and penetration depths, and the model's part predictions are not fitted to the test set. The admitted limitation that 'three mutually contacting parts cannot be represented using only two volumes' is a scope restriction on the method, not a circular step, and the failure cases in Figure 10 are disclosed empirical weaknesses rather than hidden reuse of the conclusion. Citations to prior systems such as CLAY, Trellis, Dora, and Hunyuan3D-2 are standard engineering reuse, and the few references involving the present authors (Meshtron, EdgeRunner) are related-work only and not load-bearing. No quoted step reduces to its own input, so the appropriate finding is no circularity.

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

The approach rests on empirical assumptions about object part graphs and on a hand-tuned data curation pipeline. The free parameters are the packing volume count, the occupancy filtering thresholds, and the heuristic contact and contraction settings. No new physical entities are introduced.

free parameters (4)
  • Number of packing volumes = 2
    Fixed to two volumes so that the output length is constant; this choice is motivated by the bipartite observation and directly causes the known failure for three mutually touching parts (Sec 4.5).
  • Occupancy filtering thresholds = 0.001 absolute; 0.1 min/max ratio
    These thresholds in Eq 1 discard training samples with nearly empty volumes or highly unbalanced occupancy, shaping the training distribution and the reported segmentation diversity.
  • Edge dilation amount for contact detection = One SDF voxel size
    Parts are dilated by the voxel size before collision detection so tangential contacts become graph edges; this affects which parts are considered fused and hence the packing graph (Sec 3.1).
  • Penetration-depth edge weight = Penetration depth
    The greedy contraction algorithm contracts the edge with greatest penetration depth in each odd cycle; this choice determines which semantic parts get merged and is heuristic, with no optimality guarantee (Algorithm 1).
assumptions (4)
  • domain assumption Most real part-connectivity graphs are bipartite or become bipartite after contracting a few edges.
    This motivates dual volume packing (Sec 3.1, Fig 2). The paper acknowledges in Sec 4.5 that three mutually touching parts cannot be represented, so the assumption is only approximate.
  • domain assumption Scene graph nodes, connected components, and the three heuristic merge rules yield usable ground-truth parts.
    Part-level dataset curation in Sec 3.2 relies on these rules; no external validation of semantic correctness is provided, and Sec 4.5 notes the annotations are noisy and inconsistent.
  • domain assumption Two watertight SDF volumes can faithfully encode and reconstruct the original assembled object after packing.
    The VAE operates on volumes produced by watertight conversion (Sec 3.3 and A.2); the paper reports artifacts on small parts, indicating that this assumption holds only approximately.
  • domain assumption Greedy odd-cycle contraction produces a bipartite graph with acceptable semantic loss.
    The NP-hard optimal contraction is replaced by a greedy heuristic (Sec 3.1, Algorithm 1); no bound on semantic loss is given, and Figure 10 shows failure cases where internal structures are incorrect.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient Part-level 3D Object Generation via Dual Volume Packing." pith.science (2026). https://pith.science/paper/BOH7L5FQ

@misc{pith2026250609980,
  author       = {Pith},
  title        = {Pith review of: Efficient Part-level 3D Object Generation via Dual Volume Packing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BOH7L5FQ}},
  note         = {Machine review of arXiv:2506.09980}
}
read the original abstract

Recent progress in 3D object generation has greatly improved both the quality and efficiency. However, most existing methods generate a single mesh with all parts fused together, which limits the ability to edit or manipulate individual parts. A key challenge is that different objects may have a varying number of parts. To address this, we propose a new end-to-end framework for part-level 3D object generation. Given a single input image, our method generates high-quality 3D objects with an arbitrary number of complete and semantically meaningful parts. We introduce a dual volume packing strategy that organizes all parts into two complementary volumes, allowing for the creation of complete and interleaved parts that assemble into the final object. Experiments show that our model achieves better quality, diversity, and generalization than previous image-based part-level generation methods.

Figures

Figures reproduced from arXiv: 2506.09980 by the authors.

Figure 1
Figure 1. End-to-end Part-level Image-to-3D Generation. We present a method to generate high￾quality 3D shape composed of individual and complete parts from a single-view image. Our method is trained only with 3D native information and can generate part-level meshes in about 30 seconds without relying on 2D segmentation prior models. an arbitrary number of parts within a fixed time. We identify the key challenge to be the han… view at source ↗
Figure 2
Figure 2. Dual Volume Packing. Given a 3D mesh with part-level annotations, we propose to convert the part-connectivity graph into a bipartite graph, such that all parts can be packed into two volumes. Within each volume, parts do not contact each other, thus can be separated during mesh extraction. 3 Methodology Similar to previous 3D latent denoising models [56, 22, 59, 54, 46], our model takes a single-view image as input … view at source ↗
Figure 3
Figure 3. Network Architecture. Our model takes a single-view image as the input condition, and generate the dual latents at the same time with a flow model. The latents are decoded to dual volumes, which can be divided into parts and assembled back to the whole mesh. cross-attention. Finally, a rectified flow model is trained to denoise the latent codes conditioned on these image features. VAE Model. The VAE consists of a du… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Comparison on Image-to-3D Generation. Our method generates part-level meshes with competitive quality from single-view images compared to previous methods. 4.2 Qualitative Comparisons. Our model is a latent denoising framework for image-to-3D generation, following the …
Figure 5
Figure 5. Figure 5: Comparison on Part-level 3D Generation. Our method directly generate complete parts, while other methods require mesh segmentation and part completion. Single Volume Dual Volumes Method Hunyuan3D-2 [59] Hi3DGen [54] TripoSG [22] Ours† Ours ULIP [48] 0.1609 0.1641 0.172…
Figure 6
Figure 6. Figure 6: Diversity on Image-to-3D Generation. We can generate diverse and meaningful parts with different random seeds. Input Image w/o P-VAE w/ P-VAE Input Image w/o Data Cleaning w/ Data Cleaning [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Ablation Study. We ablate different training designs and compare the generation quality. 4.4 Ablation Study We conduct ablation studies on key design choices in the part-level finetuning stage, as shown in [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Statistics on the number of part in the processed dataset. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: More Comparisons on Image-to-3D Generation. signed distance field. Using Marching Cubes [29], we extract a watertight surface from the resulting signed distance field. Following Dora [5], we sample 32768 uniformly distributed surface points and 16384 salient edge point…
Figure 10
Figure 10. Figure 10: Limitations and Failure Cases. We showcase some failure cases of our model. parts. Compared with baseline methods, our generated shapes exhibit better surface smoothness, part separation, and overall fidelity. B.2 Failure Cases In [PITH_FULL_IMAGE:figures/full_fig_p0…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 5 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. PartDiffuser: Part-wise 3D Mesh Generation via Discrete Diffusion

    cs.CV 2025-11 unverdicted novelty 7.0 of 10

    PartDiffuser is a semi-autoregressive discrete diffusion framework that generates high-fidelity 3D meshes from point clouds by combining inter-part autoregression with intra-part parallel diffusion using a part-aware ...

  2. EditVerse3D: High-Quality 3D Object Editing with Region-Aware Learning

    cs.CV 2026-07 conditional novelty 6.0 of 10

    An end-to-end 3D editing framework achieves high-fidelity local edits from coarse bounding boxes and 2D image prompts using region-aware loss reweighting and a large-scale parts-derived training dataset.

  3. AssemLM: A Spatial Reasoning Multimodal Large Language Model for Robotic Assembly

    cs.RO 2026-04 unverdicted novelty 6.0 of 10

    AssemLM uses a specialized point cloud encoder inside a multimodal LLM to reach state-of-the-art 6D pose prediction for assembly tasks, backed by a new 900K-sample benchmark called AssemBench.

  4. Twisted Fiber Bundle Codes over Group Algebras

    quant-ph 2026-04 unverdicted novelty 6.0 of 10

    Singular chain-compatible fiber twists over group algebras can increase CSS encoded dimension k at fixed blocklength n while examples keep distance d unchanged.

  5. Muses: Designing, Composing, Generating Nonexistent Fantasy 3D Creatures without Training

    cs.CV 2026-01 conditional novelty 6.0 of 10

    Muses creates new fantasy 3D animals by designing a combined skeleton, fusing voxel parts from separate 3D models along that skeleton, then restyling textures via image editing — with no training.

Reference graph

Works this paper leans on

61 extracted references · 15 canonical work pages · cited by 5 Pith papers

  1. [1]

    Large-vocabulary 3d diffusion model with transformer.arXiv preprint arXiv:2309.07920, 2023

    Ziang Cao, Fangzhou Hong, Tong Wu, Liang Pan, and Ziwei Liu. Large-vocabulary 3d diffusion model with transformer.arXiv preprint arXiv:2309.07920, 2023

    arXiv 2023

  2. [2]

    Single-stage diffusion nerf: A unified approach to 3d generation and reconstruction.arXiv preprint arXiv:2304.06714, 2023

    Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion nerf: A unified approach to 3d generation and reconstruction.arXiv preprint arXiv:2304.06714, 2023

    arXiv 2023

  3. [3]

    PrimDiffusion: Volumetric Primitives Diffusion for 3D Human Generation

    Zhaoxi Chen, Fangzhou Hong, Haiyi Mei, Guangcong Wang, Lei Yang, and Ziwei Liu. Primdif- fusion: V olumetric primitives diffusion for 3d human generation.arXiv preprint arXiv:2312.04559, 2023

    work page Pith review arXiv 2023

  4. [4]

    Partgen: Part-level 3d generation and reconstruction with multi-view diffusion models.arXiv preprint arXiv:2412.18608, 2024

    Minghao Chen, Roman Shapovalov, Iro Laina, Tom Monnier, Jianyuan Wang, David Novotny, and Andrea Vedaldi. Partgen: Part-level 3d generation and reconstruction with multi-view diffusion models.arXiv preprint arXiv:2412.18608, 2024

    arXiv 2024

  5. [5]

    Dora: Sampling and benchmarking for 3d shape variational auto-encoders.arXiv preprint arXiv:2412.17808, 2024

    Rui Chen, Jianfeng Zhang, Yixun Liang, Guan Luo, Weiyu Li, Jiarui Liu, Xiu Li, Xiaoxiao Long, Jiashi Feng, and Ping Tan. Dora: Sampling and benchmarking for 3d shape variational auto-encoders.arXiv preprint arXiv:2412.17808, 2024

    arXiv 2024

  6. [6]

    Meshxl: Neural coordinate field for generative 3d foundation models.arXiv preprint arXiv:2405.20853, 2024

    Sijin Chen, Xin Chen, Anqi Pang, Xianfang Zeng, Wei Cheng, Yijun Fu, Fukun Yin, Yanru Wang, Zhibin Wang, Chi Zhang, et al. Meshxl: Neural coordinate field for generative 3d foundation models.arXiv preprint arXiv:2405.20853, 2024

    arXiv 2024

  7. [7]

    Meshanything: Artist-created mesh generation with autoregressive transformers.arXiv preprint arXiv:2406.10163, 2024

    Yiwen Chen, Tong He, Di Huang, Weicai Ye, Sijin Chen, Jiaxiang Tang, Xin Chen, Zhongang Cai, Lei Yang, Gang Yu, et al. Meshanything: Artist-created mesh generation with autoregressive transformers.arXiv preprint arXiv:2406.10163, 2024

    arXiv 2024

  8. [8]

    Meshanything v2: Artist-created mesh generation with adjacent mesh tokenization

    Yiwen Chen, Yikai Wang, Yihao Luo, Zhengyi Wang, Zilong Chen, Jun Zhu, Chi Zhang, and Guosheng Lin. Meshanything v2: Artist-created mesh generation with adjacent mesh tokenization. arXiv preprint arXiv:2408.02555, 2024

    arXiv 2024

Show all 61 references
  1. [9]

    3dtopia-xl: Scaling high-quality 3d asset generation via primitive diffusion.arXiv preprint arXiv:2409.12957, 2024

    Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong, Ziang Cao, Fangzhou Hong, Yushi Lan, Tengfei Wang, Haozhe Xie, Tong Wu, Shunsuke Saito, et al. 3dtopia-xl: Scaling high-quality 3d asset generation via primitive diffusion.arXiv preprint arXiv:2409.12957, 2024

    2024 arXiv

  2. [10]

    Sdfusion: Multimodal 3d shape completion, reconstruction, and generation

    Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. InConference on Computer Vision and Pattern Recognition (CVPR), pages 4456–4465, 2023

    2023

  3. [11]

    Objaverse-xl: A universe of 10m+ 3d objects.arXiv preprint arXiv:2307.05663, 2023

    Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram V oleti, Samir Yitzhak Gadre, et al. Objaverse-xl: A universe of 10m+ 3d objects.arXiv preprint arXiv:2307.05663, 2023

    2023 arXiv

  4. [12]

    Objaverse: A universe of annotated 3d objects

    Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects. InConference on Computer Vision and Pattern Recognition (CVPR), pages 13142–...

    2023

  5. [13]

    3dgen: Triplane latent diffusion for textured mesh generation.arXiv preprint arXiv:2303.05371, 2023

    Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas O˘guz. 3dgen: Triplane latent diffusion for textured mesh generation.arXiv preprint arXiv:2303.05371, 2023. 10

    2023 arXiv

  6. [14]

    Meshtron: High-fidelity, artist-like 3d mesh generation at scale.arXiv preprint arXiv:2412.09548, 2024

    Zekun Hao, David W Romero, Tsung-Yi Lin, and Ming-Yu Liu. Meshtron: High-fidelity, artist-like 3d mesh generation at scale.arXiv preprint arXiv:2412.09548, 2024

    2024 arXiv

  7. [15]

    Obtaining a bipartite graph by contracting few edges.SIAM Journal on Discrete Mathematics, 27(4):2143–2156, 2013

    Pinar Heggernes, Pim Van’T Hof, Daniel Lokshtanov, and Christophe Paul. Obtaining a bipartite graph by contracting few edges.SIAM Journal on Discrete Mathematics, 27(4):2143–2156, 2013

    2013

  8. [16]

    3dtopia: Large text-to-3d generation model with hybrid diffusion priors.arXiv preprint arXiv:2403.02234, 2024

    Fangzhou Hong, Jiaxiang Tang, Ziang Cao, Min Shi, Tong Wu, Zhaoxi Chen, Tengfei Wang, Liang Pan, Dahua Lin, and Ziwei Liu. 3dtopia: Large text-to-3d generation model with hybrid diffusion priors.arXiv preprint arXiv:2403.02234, 2024

    2024 arXiv

  9. [17]

    Shap-e: Generating conditional 3d implicit functions.arXiv preprint arXiv:2305.02463, 2023

    Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions.arXiv preprint arXiv:2305.02463, 2023

    2023 arXiv

  10. [18]

    Segment anything

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015–4026, 2023

    2023

  11. [19]

    Unleashing vecset diffusion model for fast shape generation, 2025

    Zeqiang Lai, Yunfei Zhao, Zibo Zhao, Haolin Liu, Fuyun Wang, Huiwen Shi, Xianghui Yang, Qinxiang Lin, Jinwei Huang, Yuhong Liu, Jie Jiang, Chunchao Guo, and Xiangyu Yue. Unleashing vecset diffusion model for fast shape generation, 2025

    2025

  12. [20]

    Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation

    Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. arXiv preprint arXiv:2403.12019, 2024

    2024

  13. [21]

    Craftsman: High-fidelity mesh generation with 3d native generation and interactive geometry refiner.arXiv preprint arXiv:2405.14979, 2024

    Weiyu Li, Jiarui Liu, Rui Chen, Yixun Liang, Xuelin Chen, Ping Tan, and Xiaoxiao Long. Craftsman: High-fidelity mesh generation with 3d native generation and interactive geometry refiner.arXiv preprint arXiv:2405.14979, 2024

    2024 arXiv

  14. [22]

    Triposg: High-fidelity 3d shape synthesis using large-scale rectified flow models.arXiv preprint arXiv:2502.06608, 2025

    Yangguang Li, Zi-Xin Zou, Zexiang Liu, Dehu Wang, Yuan Liang, Zhipeng Yu, Xingchao Liu, Yuan-Chen Guo, Ding Liang, Wanli Ouyang, et al. Triposg: High-fidelity 3d shape synthesis using large-scale rectified flow models.arXiv preprint arXiv:2502.06608, 2025

    2025 arXiv

  15. [23]

    Treemeshgpt: Artistic mesh generation with autoregressive tree sequencing.arXiv preprint arXiv:2503.11629, 2025

    Stefan Lionar, Jiabin Liang, and Gim Hee Lee. Treemeshgpt: Artistic mesh generation with autoregressive tree sequencing.arXiv preprint arXiv:2503.11629, 2025

    2025 arXiv

  16. [24]

    Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

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

    2022 arXiv

  17. [25]

    Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003, 2022

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003, 2022

    2022 arXiv

  18. [26]

    One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization.arXiv preprint arXiv:2306.16928, 2023

    Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Zexiang Xu, Hao Su, et al. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization.arXiv preprint arXiv:2306.16928, 2023

    2023 arXiv

  19. [27]

    Meshdiffusion: Score-based generative 3d mesh modeling.arXiv preprint arXiv:2303.08133, 2023

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

    2023 arXiv

  20. [28]

    Partfield: Learning 3d feature fields for part segmentation and beyond.arXiv preprint arXiv:2504.11451, 2025

    Minghua Liu, Mikaela Angelina Uy, Donglai Xiang, Hao Su, Sanja Fidler, Nicholas Sharp, and Jun Gao. Partfield: Learning 3d feature fields for part segmentation and beyond.arXiv preprint arXiv:2504.11451, 2025

    2025 arXiv

  21. [29]

    Marching cubes: A high resolution 3d surface construction algorithm

    William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. InSeminal graphics: pioneering efforts that shaped the field, pages 347–353. 1998

    1998

  22. [30]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. InEuropean Conference on Computer Vision (ECCV), 2020. 11

    2020

  23. [31]

    Diffrf: Rendering-guided 3d radiance field diffusion

    Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. InConference on Computer Vision and Pattern Recognition (CVPR), pages 4328–4338, 2023

    2023

  24. [32]

    Point-e: A system for generating 3d point clouds from complex prompts.arXiv preprint arXiv:2212.08751, 2022

    Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts.arXiv preprint arXiv:2212.08751, 2022

    2022 arXiv

  25. [33]

    Autodecoding latent 3d diffusion models.arXiv preprint arXiv:2307.05445, 2023

    Evangelos Ntavelis, Aliaksandr Siarohin, Kyle Olszewski, Chaoyang Wang, Luc Van Gool, and Sergey Tulyakov. Autodecoding latent 3d diffusion models.arXiv preprint arXiv:2307.05445, 2023

    2023 arXiv

  26. [34]

    Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  27. [35]

    Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

    2024 arXiv

  28. [36]

    The four-colour theorem

    Neil Robertson, Daniel Sanders, Paul Seymour, and Robin Thomas. The four-colour theorem. journal of combinatorial theory, Series B, 70(1):2–44, 1997

    1997

  29. [37]

    Meshgpt: Generating triangle meshes with decoder-only transformers

    Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. InConference on Computer Vision and Pattern Recognition (CVPR), pages 19...

    2024

  30. [38]

    V olumediffusion: Flexible text-to-3d generation with efficient volumetric encoder.arXiv preprint arXiv:2312.11459, 2023

    Zhicong Tang, Shuyang Gu, Chunyu Wang, Ting Zhang, Jianmin Bao, Dong Chen, and Baining Guo. V olumediffusion: Flexible text-to-3d generation with efficient volumetric encoder.arXiv preprint arXiv:2312.11459, 2023

    2023 arXiv

  31. [39]

    Segment any mesh: Zero-shot mesh part segmentation via lifting segment anything 2 to 3d.arXiv preprint arXiv:2408.13679, 2024

    George Tang, William Zhao, Logan Ford, David Benhaim, and Paul Zhang. Segment any mesh: Zero-shot mesh part segmentation via lifting segment anything 2 to 3d.arXiv preprint arXiv:2408.13679, 2024

    2024 arXiv

  32. [40]

    Edgerunner: Auto-regressive auto-encoder for artistic mesh generation.arXiv preprint arXiv:2409.18114, 2024

    Jiaxiang Tang, Zhaoshuo Li, Zekun Hao, Xian Liu, Gang Zeng, Ming-Yu Liu, and Qinsheng Zhang. Edgerunner: Auto-regressive auto-encoder for artistic mesh generation.arXiv preprint arXiv:2409.18114, 2024

    2024 arXiv

  33. [41]

    Rodin: A generative model for sculpting 3d digital avatars using diffusion

    Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. InConference on Computer Vision and Pattern Recognition (CVPR), ...

    2023

  34. [42]

    iflame: Inter- leaving full and linear attention for efficient mesh generation.arXiv preprint arXiv:2503.16653, 2025

    Hanxiao Wang, Biao Zhang, Weize Quan, Dong-Ming Yan, and Peter Wonka. iflame: Inter- leaving full and linear attention for efficient mesh generation.arXiv preprint arXiv:2503.16653, 2025

    2025 arXiv

  35. [43]

    Pivotmesh: Generic 3d mesh generation via pivot vertices guidance.arXiv preprint arXiv:2405.16890, 2024

    Haohan Weng, Yikai Wang, Tong Zhang, CL Chen, and Jun Zhu. Pivotmesh: Generic 3d mesh generation via pivot vertices guidance.arXiv preprint arXiv:2405.16890, 2024

    2024 arXiv

  36. [44]

    Scaling mesh generation via compressive tokenization

    Haohan Weng, Zibo Zhao, Biwen Lei, Xianghui Yang, Jian Liu, Zeqiang Lai, Zhuo Chen, Yuhong Liu, Jie Jiang, Chunchao Guo, et al. Scaling mesh generation via compressive tokenization. arXiv preprint arXiv:2411.07025, 2024

    2024 arXiv

  37. [45]

    Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer.arXiv preprint arXiv:2405.14832, 2024

    Shuang Wu, Youtian Lin, Feihu Zhang, Yifei Zeng, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer.arXiv preprint arXiv:2405.14832, 2024

    2024 arXiv

  38. [46]

    Structured 3d latents for scalable and versatile 3d generation.arXiv preprint arXiv:2412.01506, 2024

    Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation.arXiv preprint arXiv:2412.01506, 2024. 12

    2024 arXiv

  39. [47]

    Brepgen: A b-rep generative diffusion model with structured latent geometry.ACM Transactions on Graphics (TOG), 43(4):1–14, 2024

    Xiang Xu, Joseph Lambourne, Pradeep Jayaraman, Zhengqing Wang, Karl Willis, and Yasutaka Furukawa. Brepgen: A b-rep generative diffusion model with structured latent geometry.ACM Transactions on Graphics (TOG), 43(4):1–14, 2024

    2024

  40. [48]

    Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding

    Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. InConference on Computer Vision and Pattern Re...

    2023

  41. [49]

    Ulip-2: Towards scalable multimodal pre-training for 3d understanding

    Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Junnan Li, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, et al. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. InConference on Computer Vision and Pattern Recognition (CVP...

    2024

  42. [50]

    An object is worth 64x64 pixels: Generating 3d object via image diffusion.arXiv preprint arXiv:2408.03178, 2024

    Xingguang Yan, Han-Hung Lee, Ziyu Wan, and Angel X Chang. An object is worth 64x64 pixels: Generating 3d object via image diffusion.arXiv preprint arXiv:2408.03178, 2024

    2024 arXiv

  43. [51]

    Sampart3d: Segment any part in 3d objects.arXiv preprint arXiv:2411.07184, 2024

    Yunhan Yang, Yukun Huang, Yuan-Chen Guo, Liangjun Lu, Xiaoyang Wu, Edmund Y Lam, Yan-Pei Cao, and Xihui Liu. Sampart3d: Segment any part in 3d objects.arXiv preprint arXiv:2411.07184, 2024

    2024 arXiv

  44. [52]

    Holopart: Generative 3d part amodal segmentation.arXiv preprint arXiv:2504.07943, 2025

    Yunhan Yang, Yuan-Chen Guo, Yukun Huang, Zi-Xin Zou, Zhipeng Yu, Yangguang Li, Yan- Pei Cao, and Xihui Liu. Holopart: Generative 3d part amodal segmentation.arXiv preprint arXiv:2504.07943, 2025

    2025 arXiv

  45. [53]

    Mosaic-sdf for 3d generative models.arXiv preprint arXiv:2312.09222, 2023

    Lior Yariv, Omri Puny, Natalia Neverova, Oran Gafni, and Yaron Lipman. Mosaic-sdf for 3d generative models.arXiv preprint arXiv:2312.09222, 2023

    2023 arXiv

  46. [54]

    Hi3dgen: High-fidelity 3d geometry generation from images via normal bridging.arXiv preprint arXiv:2503.22236, 3, 2025

    Chongjie Ye, Yushuang Wu, Ziteng Lu, Jiahao Chang, Xiaoyang Guo, Jiaqing Zhou, Hao Zhao, and Xiaoguang Han. Hi3dgen: High-fidelity 3d geometry generation from images via normal bridging.arXiv preprint arXiv:2503.22236, 3, 2025

    2025 arXiv

  47. [55]

    3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models.arXiv preprint arXiv:2301.11445, 2023

    Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models.arXiv preprint arXiv:2301.11445, 2023

    2023 arXiv

  48. [56]

    Clay: A controllable large-scale generative model for creating high-quality 3d assets.arXiv preprint arXiv:2406.13897, 2024

    Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets.arXiv preprint arXiv:2406.13897, 2024

    2024 arXiv

  49. [57]

    Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation.arXiv preprint arXiv:2306.17115, 2023

    Zibo Zhao, Wen Liu, Xin Chen, Xianfang Zeng, Rui Wang, Pei Cheng, Bin Fu, Tao Chen, Gang Yu, and Shenghua Gao. Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation.arXiv preprint arXiv:2306.17115, 2023

    2023 arXiv

  50. [58]

    Deepmesh: Auto-regressive artist-mesh creation with reinforcement learning.arXiv preprint arXiv:2503.15265, 2025

    Ruowen Zhao, Junliang Ye, Zhengyi Wang, Guangce Liu, Yiwen Chen, Yikai Wang, and Jun Zhu. Deepmesh: Auto-regressive artist-mesh creation with reinforcement learning.arXiv preprint arXiv:2503.15265, 2025

    2025 arXiv

  51. [59]

    Hunyuan3d 2.0: Scaling diffusion models for high resolution textured 3d assets generation.arXiv preprint arXiv:2501.12202, 2025

    Zibo Zhao, Zeqiang Lai, Qingxiang Lin, Yunfei Zhao, Haolin Liu, Shuhui Yang, Yifei Feng, Mingxin Yang, Sheng Zhang, Xianghui Yang, et al. Hunyuan3d 2.0: Scaling diffusion models for high resolution textured 3d assets generation.arXiv preprint arXiv:2501.12202, 2025

    2025 arXiv

  52. [60]

    Locally attentional sdf diffusion for controllable 3d shape generation.arXiv preprint arXiv:2305.04461, 42(4):1–13, 2023

    Xin-Yang Zheng, Hao Pan, Peng-Shuai Wang, Xin Tong, Yang Liu, and Heung-Yeung Shum. Locally attentional sdf diffusion for controllable 3d shape generation.arXiv preprint arXiv:2305.04461, 42(4):1–13, 2023

    2023 arXiv

  53. [61]

    Uni3d: Exploring unified 3d representation at scale

    Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3d: Exploring unified 3d representation at scale. InInternational Conference on Learning Representations (ICLR), 2024. 13 A More Implementation Details A.1 Bipartite Contraction We detail ...

    2024

Pith tools

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