Pith. sign in

REVIEW 4 major objections 6 minor 4 cited by

AutoPartGen: Autogressive 3D Part Generation and Discovery

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read AutoPartGen generates 3D objects part by part, conditioning each new part on the ones already generated, and claims the pieces assemble into a coherent object without any extra optimization step.

desk verdict Genuinely new autoregressive part-generation in a latent 3D space, with real gains on mask-guided completion, but the paper's headline 'part discovery' claim is only shown qualitatively and is not quantified. read the letter →

arxiv 2507.13346 v2 pith:CSEYX2WF submitted 2025-07-17 cs.CV

Minghao Chen , Jianyuan Wang , Roman Shapovalov , Tom Monnier , Hyunyoung Jung , Dilin Wang , Rakesh Ranjan , Iro Laina
show 1 more author
Andrea Vedaldi
This is my paper
classification cs.CV
keywords autoregressive3Dgenerationpartdecompositionlatentdiffusioncompositionalspace3DShape2VecSetcompletionimage-to-3Ddiscovery
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 introduces AutoPartGen, a model that generates 3D objects as sequences of parts rather than as one monolithic surface. Given an image, 2D part masks, or an existing 3D shell, it predicts one part at a time, conditioning each new part on the parts already generated and on an overall shape code, and stops when it emits an end-of-parts token. The central claim is that the underlying latent representation, built on the 3DShape2VecSet encoder-decoder, is compositional: concatenating the latent codes of two parts decodes to the union of their surfaces. If true, this makes it possible to assemble coherent compositional objects without any post-hoc optimization, and the paper reports that this design beats existing part-aware generators on part completion on PartObjaverse-Tiny.

What carries the argument

The load-bearing object is the compositional VecSet latent code. In the 3DShape2VecSet representation, a 3D surface is encoded into transformer tokens and decoded, at each query point, into a signed distance value; the paper observes that concatenating two such token sequences yields a new code whose decoded surface is the union of the two parts. That property converts part generation into an autoregressive sequence problem: at step $k$, a transformer-based diffusion model samples $z^{(k)}$ from $p(z^{(k)} | \tilde{z}, z^{(1..k-1)}, y)$, where $\tilde{z}$ pins down the overall object, $z^{(1..k-1)}$ are the codes of already-generated parts (fused by re-encoding their decoded union), and $y$ is either nothing, an image, or an image plus a 2D part mask. An end-of-parts empty-shape token stops the sequence.

What would settle it

Take a PartObjaverse-Tiny object with at least three parts, encode each part separately, concatenate the codes in generation order, decode, and compare the resulting surface with the ground-truth union. If part-level IoU drops noticeably when moving from two to five or ten concatenated parts, or if seams between decoded parts become visible, the no-optimization assembly claim fails.

Watch

Extended reading notes

Core claim

The paper's central discovery is empirical: the VecSet latent space, which encodes a point cloud into a small set of tokens and decodes a signed distance field from those tokens, supports direct concatenation. A latent formed as $z^{(1)} \oplus z^{(2)}$ decodes into a surface that approximates $x^{(1)} \cup x^{(2)}$, with no retraining, and the paper builds an entire autoregressive generation scheme on this property. A single diffusion transformer is trained to sample the next part code conditioned on previously generated part codes, an overall object code, and optionally an input image and masked part image; the same model handles object-to-parts, image-to-parts, and masks-to-parts. The model also outputs a special empty-shape token to decide when the part sequence is complete, so the number and type of parts is discovered rather than fixed. On PartObjaverse-Tiny part completion, it reports part IoU 0.665, F-score 0.861, and Chamfer distance 0.047, exceeding the published numbers of HoloPart and PartGen on the same filtered evaluation.

Load-bearing premise

Everything rests on the assumption that concatenating latent codes keeps decoding to the union of the parts reliably, not just for the two-part illustration but for long autoregressive sequences of many generated parts.

Editorial extensions

If this is right

  • One unified model can decompose an existing 3D object, reconstruct parts from a single image, and follow user-provided 2D masks to control granularity.
  • Because the next part is conditioned on previously generated parts, the model avoids overlapping and intersecting parts that independent part samplers produce, as shown in the ablation study.
  • Generated parts assemble into a coherent object without any optimization stage, which simplifies pipelines for editable 3D assets.
  • When combined with a text-to-image generator, the same decomposition mechanism extends to whole scenes and city-scale tile generation.

Reading between the lines

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

  • If the concatenation property degrades with code length, the method would need a fusion or refinement step; a direct test is to measure union reconstruction error for 3, 5, and 10 concatenated part codes.
  • The autoregressive ordering and the end-of-parts stopping rule are learned from artist-made mesh hierarchies, so the model inherits a particular convention for what counts as a part; on objects outside that convention, part discovery may differ from a user's expectation.
  • The scene and city demonstrations stitch together repeated applications of the same model; training the model directly on unbounded scenes would be a separate step, and the paper itself notes the latent space currently bounds scene size.
  • The compositionality observation could transfer to other token-based 3D autoencoders: if any VecSet-style latent has the union-concatenation property, single-model part generation may become a general recipe rather than a property of this one checkpoint.
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 / 6 minor

Summary. The paper introduces AutoPartGen, an autoregressive latent-diffusion model for 3D part generation and discovery. The method builds on the 3DShape2VecSet latent representation and claims that concatenating part codes decodes to the union of the corresponding surfaces. A single DiT-based diffusion model generates one part at a time, conditioned on previously generated parts, an overall object code, and, optionally, an input image and part masks; generation stops via a learned [EoT] token. The authors evaluate mask-controlled part completion on PartObjaverse-Tiny against HoloPart and PartGen, report an autoregressive ablation, and show qualitative results for image-to-parts and object-to-parts scenarios, as well as scene and city generation.

Significance. If the central claims hold, AutoPartGen would be a useful step toward unified, controllable 3D part generation: it avoids multi-view image generators and external 3D segmentation models, handles three input modalities with one model, and its learned stopping criterion addresses the variable-number-of-parts problem. The autoregressive ablation (Table 2) provides positive evidence for the key design choice, and the large-scale training pipeline described in Section 4.1 is itself a contribution. However, the quantitative evidence is currently limited to the mask-controlled setting, and the compositionality property that underpins the method is not quantitatively validated; these gaps must be closed before the state-of-the-art claim can be accepted.

major comments (4)
  1. [§4.3, Table 1] The only quantitative evaluation is the mask-controlled part-completion setting, and Section 4.3 states "We focus on mask-controlled part generation." The image-to-parts and object-to-parts scenarios, which are central to the abstract's claim of discovering parts without external segmentation models, are shown only qualitatively (Figs. 4 and 5 for AutoPartGen; App. Fig. 9 for the PartGen comparison). Because the mask-controlled setting provides the part decomposition as an input, it does not test the discovery capability that distinguishes AutoPartGen from HoloPart. To support the state-of-the-art claim, the authors should report quantitative metrics (e.g., part IoU, part-count accuracy, or segmentation coverage) for at least one discovery benchmark, such as object-to-parts decomposition on Google Scanned Objects or a held-out part-annotated dataset.
  2. [§3.1, Fig. 3] The paper's entire no-optimization assembly mechanism rests on the claim that concatenation of two VecSet codes decodes to the union of the corresponding surfaces (Section 3.1: "the tokens can be concatenated to form a new latent vector z = z(1) ⊕ z(2) that decodes into a new surface x = x(1) ∪ x(2)"). This property is verified only by a single illustrative figure. No quantitative measurement of union reconstruction error is provided, and the paper does not test whether the property degrades as the number of concatenated codes grows or as part scales vary. Since Eq. (1) and the fusion step z(1,...,k−1) = E(∪ sample_N D(·|z(j))) both depend on this property, please add a quantitative evaluation of concatenation fidelity as a function of part count and compare with alternative fusion operations (e.g., averaging or additional cross-attention).
  3. [Table 1, Table 3] The main quantitative result (Table 1) reports gains that are modest in absolute terms (e.g., part IoU 0.665 vs 0.658 for HoloPart, CD 0.047 vs 0.065), yet no standard errors or repeated runs are reported. Additionally, the default guidance scales w_img=7, w_geom=4 are selected using Table 3 on the same evaluation set, so the comparison may be optimistic due to test-set overfitting. Please provide variance estimates (multiple seeds or evaluation runs) and select hyperparameters on a validation split, or at least justify that the chosen scales do not advantage AutoPartGen relative to the baselines, which have their own hyperparameters.
  4. [§4.3, Table 1] The comparison protocol gives different input modalities to the three methods: HoloPart receives the partial 3D object and a 3D part segmentation, PartGen receives four masked views, and AutoPartGen receives the partial 3D mesh plus 2D part masks. Because AutoPartGen is the only method that receives both geometric and 2D mask information, it is unclear how much of the improvement is due to the autoregressive design rather than the additional geometry condition. Please include an ablation in which AutoPartGen is run with the same input types as HoloPart or PartGen (or explicitly discuss why the adapted protocol is the fairest apples-to-apples comparison).
minor comments (6)
  1. [§4.3] The metric description says "IoU is calculated on 643 voxel grids", which is presumably a typo for 64^3 voxel grids; please fix the exponent.
  2. [Supplementary E] The supplementary text cites "SynCity [36]" but SynCity is reference [12] in the main paper; reference [36] corresponds to SyncDreamer. Please correct the citation.
  3. [Fig. 8] The subfigure labels and caption are confusing: the caption reads "(a) Without autoregressive generation... (b) Increasing image guidance..." while the figure overlay includes "AR (b)(a) Non-AR". Please clarify which panel corresponds to which condition and ensure the caption matches the panel labels.
  4. [§3.3, Eq. (2)] The notation for the unconditional and condition-dropped velocity fields in Eq. (2) is not fully defined; in particular, "v(t,z_t,∅)" and the dummy-token replacements are described only in prose. Please spell out how each term is obtained during inference.
  5. [Table 2] The autoregressive ablation in Table 2 is trained for only 200 epochs, whereas the main model is trained for 500K iterations; please state whether the qualitative AR advantage and the magnitude of the quantitative gap persist for the full training budget.
  6. [General] The paper does not state whether the code or trained models will be released; for reproducibility, please add an availability statement.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the compositionality premise is empirical, baselines are external, and the SOTA claim is narrower than advertised but not derived from its own inputs.

full rationale

This paper is an empirical systems paper rather than a derivation chain whose conclusions are forced by its premises. The autoregressive pipeline rests on the claimed compositionality of the 3DShape2VecSet latent space ('the tokens can be concatenated to form a new latent vector z = z(1) ⊕ z(2) that decodes into a new surface x = x(1) ∪ x(2) ... without any retraining', Section 3.1, Fig. 3), but this is presented as an observed, empirically discovered property of a pretrained representation, not as a theorem derived from the paper's own equations; it is therefore not circular, though it is only qualitatively demonstrated and would benefit from quantitative union-reconstruction error measurement. The quantitative SOTA claim is anchored to Table 1 on the external PartObjaverse-Tiny benchmark, against the external HoloPart baseline and the self-cited PartGen baseline; the PartGen citation is used for comparison and data-pipeline inspiration, not as a load-bearing premise that replaces evidence. Choosing guidance scales on the same evaluation set (Table 3) is a test-set hyperparameter selection concern that weakens the strength of the SOTA claim, but it is not a fitted parameter renamed as a prediction and does not make the reported part-completion metrics equal to the selected hyperparameters by construction. The paper's own supplementary and limitations sections admit that the object-to-parts and image-to-parts discovery scenarios are only qualitatively evaluated and that granularity control is limited, which narrows the advertised scope of the claim but is a completeness/validity issue rather than circularity. No equation in the paper reduces to its own input, and no prediction is statistically forced by a fit to the same quantity.

Assumptions & free parameters 2 free parameters · 4 assumptions · 1 invented entities

The method rests on a set of assumptions: that artist-defined mesh hierarchies provide meaningful part supervision, that the VecSet latent concatenation behaves compositionally, and that standard latent diffusion generalizes to sequential part conditioning. The only hand-tuned parameters that affect the headline numbers are the two CFG guidance scales, selected on the same benchmark used for the final comparison.

free parameters (2)
  • image guidance weight w_img = 7
    Chosen as default after sweeping in Table 3 on PartObjaverse-Tiny; affects part/mask adherence and final metrics.
  • geometry guidance weight w_geom = 4
    Chosen with w_img via Table 3; higher values bias toward learned part prior, lower toward image.
assumptions (4)
  • domain assumption A 3D object is a finite disjoint union of closed regular surfaces, each with a well-defined SDF.
    Section 3, footnote 1; needed for the VecSet SDF decoding and union-of-parts semantics.
  • ad hoc to paper The concatenation of two VecSet latent codes decodes to the union of their surfaces.
    Section 3.1 and Fig. 3; stated as an empirical discovery with a single qualitative example, no quantitative validation.
  • domain assumption Mesh hierarchies in glTF/GLB training assets define semantically meaningful parts.
    Section 4.1; the model learns part definitions from artist-created mesh groupings, so part semantics are whatever those groupings encode.
  • standard math Flow matching / latent diffusion trained on this representation yields a correct conditional distribution p(z|y).
    Section 3.2; standard framework from prior work, assumed to generalize to sequential part conditioning.
invented entities (1)
  • [EoT] empty-shape token
    purpose: Signals the autoregressive model to stop generating parts, so the number of parts is determined automatically.
    A model-internal construct; the paper shows qualitative behavior but no quantitative stopping reliability metric, so there is no evidence outside the paper that the token reliably terminates generation at the correct point.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AutoPartGen: Autogressive 3D Part Generation and Discovery." pith.science (2026). https://pith.science/paper/CSEYX2WF

@misc{pith2026250713346,
  author       = {Pith},
  title        = {Pith review of: AutoPartGen: Autogressive 3D Part Generation and Discovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CSEYX2WF}},
  note         = {Machine review of arXiv:2507.13346}
}
read the original abstract

We introduce AutoPartGen, a model that generates objects composed of 3D parts in an autoregressive manner. This model can take as input an image of an object, 2D masks of the object's parts, or an existing 3D object, and generate a corresponding compositional 3D reconstruction. Our approach builds upon 3DShape2VecSet, a recent latent 3D representation with powerful geometric expressiveness. We observe that this latent space exhibits strong compositional properties, making it particularly well-suited for part-based generation tasks. Specifically, AutoPartGen generates object parts autoregressively, predicting one part at a time while conditioning on previously generated parts and additional inputs, such as 2D images, masks, or 3D objects. This process continues until the model decides that all parts have been generated, thus determining automatically the type and number of parts. The resulting parts can be seamlessly assembled into coherent objects or scenes without requiring additional optimization. We evaluate both the overall 3D generation capabilities and the part-level generation quality of AutoPartGen, demonstrating that it achieves state-of-the-art performance in 3D part generation.

Figures

Figures reproduced from arXiv: 2507.13346 by the authors.

Figure 1
Figure 1. AutoPartGen can be applied, by itself or in combination with other models, to the generation of compositional 3D objects, scenes and cities starting from 3D models, images or text. Abstract We introduce AutoPartGen, a model that generates objects composed of 3D parts in an autoregressive manner. This model can take as input an image of an object, 2D masks of the object’s parts, or an existing 3D object, and generate… view at source ↗
Figure 2
Figure 2. AutoPartGen generates parts autoregressively. At each step, a 3D latent diffusion model generate the next part, conditioned on the previously generated parts z (1,...,k) , the overall object z˜, and, optionally, an image I of the object and an image J (k) of the part. The latent representation uses 3DShape2VecSet and the diffusion model is a DiT. x (1) , . . . , x (K) , each of which is also a surface. Concretely, x… view at source ↗
Figure 3
Figure 3. Compositionality of the VecSet space. Concatenation of two latents will re￾sult in a spatial combined mesh. The intuition behind this representation is that each token vector zi encodes a local region of the surface centered at the point pi . However, the transformer allows tokens to communicate globally, which makes this interpretation somewhat loose. Empirically, we have discovered that locality, or at least compo… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Image-to-parts scenario. Given an input image, AutoPartGen recovers a compositional 3D object made up of several meaningful and complete parts. Prior works address this issue by explicitly telling the 3D reconstruction model which part to extract next. PartGen does so …
Figure 5
Figure 5. Figure 5: Object-to-parts scenario. Given an input 3D object, AutoPartGen regenerates it as a composition of meaningful and complete 3D parts. Input Mask 1 Mask 2 Input Mask 1 Mask 2 [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Masks-to-parts scenario. AutoPartGen reconstructs a compositional 3D object guided by user-provided 2D part masks. Varying these masks yields different decompositions, potentially at different levels of granularity. 4.2 Object, Image and Masks to 3D Parts Generation We…
Figure 7
Figure 7. Figure 7: Visual comparison of different completion methods. Our approach achieves better geometric coherence by considering previously generated parts in context. (a) Non-AR AR (b) [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 9
Figure 9. Figure 9: Comparison between AutoPartGen and PartGen. AutoPartGen produces more accurate geometry, as highlighted in the red circle. Additionally, its autoregressive generation prevents the overlapping parts observed in PartGen, as shown in the yellow circle. A.3 Evaluation The …
Figure 10
Figure 10. Figure 10: 3D scene generation. AutoPartGen generates 3D scenes while decomposing them into their constituent elements. The input images are generated by a 2D text-to-image generator. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: City Generation. We showcase AutoPartGen on larger scenes by integrating it within Syncity [12]. From top to bottom, the images depict a medieval town, a cozy town, and a solarpunk city. F Failure Cases Input Masked Image Part Object [PITH_FULL_IMAGE:figures/full_fig…
Figure 12
Figure 12. Figure 12: Failure Case. When there are identical parts, the model sometimes will try to predict the parts together even if masks are given. We present a failure case in [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 4 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. VoxHammer: Training-Free Precise and Coherent 3D Editing in Native 3D Space

    cs.CV 2025-08 conditional novelty 7.0 of 10

    A training-free 3D editing method that inverts a source asset into TRELLIS latent space and replaces latents plus attention K/V tokens in unedited regions during re-denosing.

  3. EditFlow3D: Automated Local Editing of 3D Assets with Trajectory Preservation

    cs.CV 2026-08 conditional novelty 6.0 of 10

    Mask-guided differential flow with a soft preservation loss enables training-free local 3D editing that keeps unedited regions close to the source asset.

  4. 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

67 extracted references · 42 canonical work pages · cited by 4 Pith papers

  1. [1]

    SPAGHETTI: editing implicit shapes through part aware generation

    Hertz Amir, Perel Or, Giryes Raja, Sorkine-Hornung Olga, and Cohen-Or Daniel. SPAGHETTI: editing implicit shapes through part aware generation. InACM Transactions on Graphics, 2022

    work page 2022

  2. [2]

    Henriques, Andrea Vedaldi, and Andrew Zisserman

    Yash Sanjay Bhalgat, Iro Laina, Joao F. Henriques, Andrea Vedaldi, and Andrew Zisserman. Contrastive Lift: 3D object instance segmentation by slow-fast contrastive fusion. InProceed- ings of Advances in Neural Information Processing Systems (NeurIPS), 2023

    work page 2023

  3. [3]

    Henriques, Andrew Zisserman, and Andrea Vedaldi

    Yash Sanjay Bhalgat, Iro Laina, Joao F. Henriques, Andrew Zisserman, and Andrea Vedaldi. N2F2: Hierarchical scene understanding with nested neural feature fields. InProceedings of the European Conference on Computer Vision (ECCV), 2024

    work page 2024

  4. [4]

    Neural part priors: Learning to optimize part-based object completion in rgb-d scans

    Aleksei Bokhovkin and Angela Dai. Neural part priors: Learning to optimize part-based object completion in rgb-d scans. InProc. CVPR, 2023

    work page 2023

  5. [5]

    PartGen: Part-level 3d generation and reconstruction with multi-view diffusion models

    Minghao Chen, Roman Shapovalov, Iro Laina, Jianyuan Wang Tom Monnier, David Novotny, and Andrea Vedaldi. PartGen: Part-level 3d generation and reconstruction with multi-view diffusion models. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    work page 2025

  6. [6]

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

    arXiv 2024

  7. [7]

    Grounded 3d-llm with referent tokens.arXiv preprint arXiv:2405.10370, 2024

    Yilun Chen, Shuai Yang, Haifeng Huang, Tai Wang, Runsen Xu, Ruiyuan Lyu, Dahua Lin, and Jiangmiao Pang. Grounded 3d-llm with referent tokens.arXiv preprint arXiv:2405.10370, 2024

    arXiv 2024

  8. [8]

    Comboverse: Compositional 3d assets creation using spatially-aware diffusion guidance

    Yongwei Chen, Tengfei Wang, Tong Wu, Xingang Pan, Kui Jia, and Ziwei Liu. Comboverse: Compositional 3d assets creation using spatially-aware diffusion guidance. InProc. ECCV, 2024

    work page 2024

Show all 67 references
  1. [9]

    Lucid- Dreamer: Domain-free generation of 3d gaussian splatting scenes

    Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Lucid- Dreamer: Domain-free generation of 3d gaussian splatting scenes. InarXiv, 2023

    2023

  2. [10]

    DetailGen3D: generative 3D geometry enhancement via data-dependent flow.arXiv, 2411.16820, 2024

    Ken Deng, Yuanchen Guo, Jingxiang Sun, Zixin Zou, Yangguang Li, Xin Cai, Yanpei Cao, Yebin Liu, and Ding Liang. DetailGen3D: generative 3D geometry enhancement via data-dependent flow.arXiv, 2411.16820, 2024

    2024 arXiv

  3. [11]

    McHugh, and Vincent Vanhoucke

    Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Rey- mann, Thomas B. McHugh, and Vincent Vanhoucke. Google Scanned Objects: A high-quality dataset of 3D scanned household items. InProc. ICRA, 2022

    2022

  4. [12]

    SynCity: training-free generation of 3d worlds.arXiv, 2503.16420, 2025

    Paul Engstler, Aleksandar Shtedritski, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. SynCity: training-free generation of 3d worlds.arXiv, 2503.16420, 2025

    2025 arXiv

  5. [13]

    GET3D: A generative model of high quality 3D textured shapes learned from images.arXiv.cs, abs/2209.11163, 2022

    Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. GET3D: A generative model of high quality 3D textured shapes learned from images.arXiv.cs, abs/2209.11163, 2022

    2022 arXiv

  6. [14]

    Murphy, and Tim Salimans

    Ruiqi Gao, Emiel Hoogeboom, Jonathan Heek, Valentin De Bortoli, Kevin P. Murphy, and Tim Salimans. Diffusion meets flow matching: Two sides of the same coin, 2024

    2024

  7. [15]

    Funkhouser

    Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas A. Funkhouser. Local deep implicit functions for 3D shape. InProc. CVPR, 2020

    2020

  8. [16]

    Freeman, and Thomas Funkhouser

    Kyle Genova, Forrester Cole, Daniel Vlasic, Aaron Sarna, William T. Freeman, and Thomas Funkhouser. Learning shape templates with structured implicit functions. InProc. CVPR, 2019

    2019

  9. [17]

    Vfusion3d: Learning scalable 3d generative models from video diffusion models

    Junlin Han, Filippos Kokkinos, and Philip Torr. Vfusion3d: Learning scalable 3d generative models from video diffusion models. InEuropean Conference on Computer Vision, pages 333–350. Springer, 2024

    2024

  10. [18]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. InProc. NeurIPS, 2021. 10

    2021

  11. [19]

    LRM: Large reconstruction model for single image to 3D

    Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. InProc. ICLR, 2024

    2024

  12. [20]

    Neural template: Topology-aware reconstruction and disentangled generation of 3d meshes

    Ka-Hei Hui, Ruihui Li, Jingyu Hu, and Chi-Wing Fu. Neural template: Topology-aware reconstruction and disentangled generation of 3d meshes. InProc. CVPR, 2022

    2022

  13. [21]

    Hénaff, Matthew M

    Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier J. Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, and João Carreira. Perceiver IO: A general...

    2022

  14. [22]

    UpFusion: novel view diffusion from unposed sparse view observations.arXiv, 2024

    Bharath Raj Nagoor Kani, Hsin-Ying Lee, Sergey Tulyakov, and Shubham Tulsiani. UpFusion: novel view diffusion from unposed sparse view observations.arXiv, 2024

    2024

  15. [23]

    3D Gaussian Splatting for real-time radiance field rendering.Proc

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3D Gaussian Splatting for real-time radiance field rendering.Proc. SIGGRAPH, 42(4), 2023

    2023

  16. [24]

    LERF: language embedded radiance fields

    Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. LERF: language embedded radiance fields. InProc. ICCV, 2023

    2023

  17. [25]

    Garfield: Group anything with radiance fields.arXiv.cs, abs/2401.09419, 2024

    Chung Min Kim, Mingxuan Wu, Justin Kerr, Ken Goldberg, Matthew Tancik, and Angjoo Kanazawa. Garfield: Group anything with radiance fields.arXiv.cs, abs/2401.09419, 2024

    2024 arXiv

  18. [26]

    Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick. Segment anything. InProc. CVPR, 2023

    2023

  19. [27]

    Decomposing NeRF for editing via feature field distillation

    Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing NeRF for editing via feature field distillation. InProc. NeurIPS, 2022

    2022

  20. [28]

    SALAD: part-level latent diffusion for 3D shape generation and manipulation

    Juil Koo, Seungwoo Yoo, Minh Hieu Nguyen, and Minhyuk Sung. SALAD: part-level latent diffusion for 3D shape generation and manipulation. InProc. ICCV, 2023

    2023

  21. [29]

    CraftsMan: high-fidelity mesh generation with 3d native generation and interactive geometry refiner.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, 2405.14979, 2024

    2024 arXiv

  22. [30]

    TripoSG: high-fidelity 3D shape synthesis using large-scale rectified flow models.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, and Yan-Pei Cao. TripoSG: high-fidelity 3D shape synthesis using large-scale rectified flow models.arXiv, 2502.06608, 2025

    2025 arXiv

  23. [31]

    Guibas, and Paul Guerrero

    Connor Lin, Niloy Mitra, Gordon Wetzstein, Leonidas J. Guibas, and Paul Guerrero. NeuForm: adaptive overfitting for neural shape editing. InProc. NeurIPS, 2022

    2022

  24. [32]

    Common diffusion noise schedules and sample steps are flawed.arXiv.cs, abs/2305.08891, 2023

    Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed.arXiv.cs, abs/2305.08891, 2023

    2023 arXiv

  25. [33]

    Part123: Part-aware 3d reconstruction from a single-view image.arXiv, 2405.16888, 2024

    Anran Liu, Cheng Lin, Yuan Liu, Xiaoxiao Long, Zhiyang Dou, Hao-Xiang Guo, Ping Luo, and Wenping Wang. Part123: Part-aware 3d reconstruction from a single-view image.arXiv, 2405.16888, 2024

    2024 arXiv

  26. [34]

    PartSLIP: low-shot part segmentation for 3D point clouds via pretrained image-language models

    Minghua Liu, Yinhao Zhu, Hong Cai, Shizhong Han, Zhan Ling, Fatih Porikli, and Hao Su. PartSLIP: low-shot part segmentation for 3D point clouds via pretrained image-language models. InProc. CVPR, 2023

    2023

  27. [35]

    Zero-1-to-3: Zero-shot one image to 3D object

    Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl V ondrick. Zero-1-to-3: Zero-shot one image to 3D object. InProc. ICCV, 2023

    2023

  28. [36]

    SyncDreamer: Generating multiview-consistent images from a single-view image.arXiv, 2309.03453, 2023

    Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. SyncDreamer: Generating multiview-consistent images from a single-view image.arXiv, 2309.03453, 2023

    2023 arXiv

  29. [37]

    Differentiable blocks world: Qualitative 3d decomposition by rendering primitives

    Tom Monnier, Jake Austin, Angjoo Kanazawa, Alexei Efros, and Mathieu Aubry. Differentiable blocks world: Qualitative 3d decomposition by rendering primitives. InProc. NeurIPS, 2023. 11

    2023

  30. [38]

    DiffFacto: controllable part-based 3D point cloud generation with cross diffusion

    George Kiyohiro Nakayama, Mikaela Angelina Uy, Jiahui Huang, Shi-Min Hu, Ke Li, and Leonidas Guibas. DiffFacto: controllable part-based 3D point cloud generation with cross diffusion. InProc. ICCV, 2023

    2023

  31. [39]

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rab...

    2024

  32. [40]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023

    2023

  33. [41]

    Barron, and Ben Mildenhall

    Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. InProc. ICLR, 2023

    2023

  34. [42]

    LangSplat: 3D language Gaussian splatting

    Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. LangSplat: 3D language Gaussian splatting. InProc. CVPR, 2024

    2024

  35. [43]

    NeRF for outdoor scene relighting

    Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. NeRF for outdoor scene relighting. InProc. ECCV, 2021

    2021

  36. [44]

    MVDream: Multi-view diffusion for 3D generation

    Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3D generation. InProc. ICLR, 2024

    2024

  37. [45]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In Proc. ICLR, 2021

    2021

  38. [46]

    DreamCraft3D: Hierarchical 3D generation with bootstrapped diffusion prior.arXiv.cs, abs/2310.16818, 2023

    Jingxiang Sun, Bo Zhang, Ruizhi Shao, Lizhen Wang, Wen Liu, Zhenda Xie, and Yebin Liu. DreamCraft3D: Hierarchical 3D generation with bootstrapped diffusion prior.arXiv.cs, abs/2310.16818, 2023

    2023 arXiv

  39. [47]

    Splatter Image: Ultra-fast single-view 3D reconstruction

    Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Splatter Image: Ultra-fast single-view 3D reconstruction. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  40. [48]

    PartSDF: part- based implicit neural representation for composite 3d shape parametrization and optimization

    Nicolas Talabot, Olivier Clerc, Arda Cinar Demirtas, Doruk Oner, and Pascal Fua. PartSDF: part- based implicit neural representation for composite 3d shape parametrization and optimization. arXiv, 2502.12985, 2025

    2025

  41. [49]

    DreamGaussian: Generative gaussian splatting for efficient 3D content creation.arXiv, 2309.16653, 2023

    Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. DreamGaussian: Generative gaussian splatting for efficient 3D content creation.arXiv, 2309.16653, 2023

    2023 arXiv

  42. [50]

    Emiris, Yannis Avrithis, and Leonidas J

    Konstantinos Tertikas, Despoina Paschalidou, Boxiao Pan, Jeong Joon Park, Mikaela Angelina Uy, Ioannis Z. Emiris, Yannis Avrithis, and Leonidas J. Guibas. PartNeRF: Generating part- aware editable 3D shapes without 3D supervision.arXiv.cs, abs/2303.09554, 2023

    2023 arXiv

  43. [51]

    Neural Feature Fusion Fields: 3D distillation of self-supervised 2D image representation

    Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural Feature Fusion Fields: 3D distillation of self-supervised 2D image representation. InProceedings of the International Conference on 3D Vision (3DV), 2022

    2022

  44. [52]

    Pro- lificDreamer: High-fidelity and diverse text-to-3D generation with variational score distillation

    Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Pro- lificDreamer: High-fidelity and diverse text-to-3D generation with variational score distillation. arXiv.cs, abs/2305.16213, 2023

    2023 arXiv

  45. [53]

    Novel view synthesis with diffusion models

    Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. InProc. ICLR, 2023

    2023

  46. [54]

    MeshLRM: large reconstruction model for high-quality mesh.arXiv, 2404.12385, 2024

    Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. MeshLRM: large reconstruction model for high-quality mesh.arXiv, 2404.12385, 2024. 12

    2024 arXiv

  47. [55]

    In- stantMesh: efficient 3D mesh generation from a single image with sparse-view large reconstruc- tion models.arXiv, 2404.07191, 2024

    Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. In- stantMesh: efficient 3D mesh generation from a single image with sparse-view large reconstruc- tion models.arXiv, 2404.07191, 2024

    2024 arXiv

  48. [56]

    GRM: Large gaussian reconstruction model for efficient 3D reconstruction and generation.arXiv, 2403.14621, 2024

    Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. GRM: Large gaussian reconstruction model for efficient 3D reconstruction and generation.arXiv, 2403.14621, 2024

    2024 arXiv

  49. [57]

    Pandora3D: A comprehensive framework for high-quality 3D shape and texture generation.arXiv, 2502.14247, 2025

    Jiayu Yang, Taizhang Shang, Weixuan Sun, Xibin Song, Ziang Chen, Senbo Wang, Shenzhou Chen, Weizhe Liu, Hongdong Li, and Pan Ji. Pandora3D: A comprehensive framework for high-quality 3D shape and texture generation.arXiv, 2502.14247, 2025

    2025 arXiv

  50. [58]

    HoloPart: generative 3d part amodal segmentation.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, 2504.07943, 2025

    2025 arXiv

  51. [59]

    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

  52. [60]

    Hi3DGen: High-fidelity 3D geometry generation from images via normal bridging.arXiv, 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, 2025

    2025

  53. [61]

    Omniseg3d: Omniversal 3d segmentation via hierarchical contrastive learning

    Haiyang Ying, Yixuan Yin, Jinzhi Zhang, Fan Wang, Tao Yu, Ruqi Huang, and Lu Fang. Omniseg3d: Omniversal 3d segmentation via hierarchical contrastive learning. InProc. CVPR, 2024

    2024

  54. [62]

    3DShape2VecSet: A 3D shape representation for neural fields and generative diffusion models

    Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3DShape2VecSet: A 3D shape representation for neural fields and generative diffusion models. InACM Transactions on Graphics, 2023

    2023

  55. [63]

    CLAY: A controllable large-scale generative model for creating high-quality 3D assets.arXiv, 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, 2024

    2024

  56. [64]

    Michelangelo: Conditional 3D shape generation based on shape-image-text aligned latent representation

    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. InProc. NeurIPS, 2023

    2023

  57. [65]

    Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J. Davison. In-place scene labelling and understanding with implicit scene representation. InProc. ICCV, 2021

    2021

  58. [66]

    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. InProc. ICLR, 2024

    2024

  59. [67]

    an isometric view of an office

    Yuchen Zhou, Jiayuan Gu, Xuanlin Li, Minghua Liu, Yunhao Fang, and Hao Su. PartSLIP++: enhancing low-shot 3d part segmentation via multi-view instance segmentation and maximum likelihood estimation.arXiv, 2312.03015, 2023. 13 Supplementary Material This supplementary material ...

    2023 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.