Pith. sign in

REVIEW 4 major objections 8 minor 7 cited by

StructureNet: Hierarchical Graph Networks for 3D Shape Generation

T0 review · 4 major / 8 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read The paper introduces StructureNet, a hierarchical graph-network autoencoder that encodes both part geometry and part hierarchy in one 256-D latent space, enabling generation, interpolation, abstraction, and editing.

desk verdict A real advance in structured shape generation with a clear architecture and strong qualitative results, but the quantitative evaluation and the sibling-only relationship assumption need scrutiny before the claims are fully trusted. read the letter →

arxiv 1908.00575 v1 pith:IELKIP5D submitted 2019-08-01 cs.GR cs.CGcs.CV

Kaichun Mo , Paul Guerrero , Li Yi , Hao Su , Peter Wonka , Niloy Mitra , Leonidas J. Guibas This is my paper · ORCID
classification cs.GRcs.CGcs.CV
keywords shapeanalysisandsynthesisgraphneuralnetworksobjectstructureautoencodergenerativemodelshierarchicalrepresentation3Dgenerationlatentspaceinterpolation
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

StructureNet is a variational autoencoder that represents a 3D shape as a hierarchy of n-ary graphs: each node is a part or part assembly, and additional edges between sibling parts record geometric relationships such as adjacency and symmetry. The central claim is that this n-ary representation, unlike methods that first force part hierarchies into binary trees, can be trained directly and stably on large shape collections, and that the resulting 256-dimensional latent space captures both continuous geometry and discrete structural variation. If true, one learned space can generate novel structured shapes, interpolate between shapes through small structural steps, recover structure from unannotated images or point clouds, and edit parts while preserving overall plausibility. The paper supports the claim with reconstruction, generation, interpolation, abstraction, and editing experiments on thousands of chairs, tables, and cabinets, where the comparison method degrades on datasets of this size.

What carries the argument

The load-bearing object is the hierarchy of n-ary graphs: a part tree with horizontal relationship edges between siblings, where a parent may have any number of children and paths to leaves need not be balanced. The machinery that carries the argument is a hierarchical variational autoencoder built from recursive graph encoders and decoders. The graph encoder performs message passing along sibling edges with a graph-isomorphism-style update, then max-pools over children to form the parent feature; the graph decoder reverses this by predicting up to $n_p = 10$ child parts and all their pairwise edges with existence probabilities, discarding parts and edges below a threshold, and then applying message passing so parts coordinate geometry. Order invariance—essential because sibling order is meaningless—comes from symmetric aggregation in the encoder and from solving a linear assignment to establish correspondences for the reconstruction loss. The structure consistency loss is the final pillar: it penalizes decoded geometries that violate decoded symmetry or adjacency edges, and by applying the relationship to whole subtrees it also enforces relationships between non-leaf parts.

What would settle it

Measure the structure-consistency error on annotated relationships that connect parts in different subtrees of the hierarchy; if that error stays large while sibling-relationship error shrinks, the sibling-only sparsification is the bottleneck.

Watch

Extended reading notes

Core claim

StructureNet's central discovery is that the structure of a 3D shape can be encoded as a hierarchy of n-ary graphs—part trees augmented by relationship edges between siblings, with edge types for adjacency, reflective symmetry, rotational symmetry, and translational symmetry—and that a recursive graph-network autoencoder over this representation yields a latent space where discrete structure and continuous geometry vary jointly and smoothly. The encoder is order-invariant: it aggregates sibling features by message passing followed by max-pooling, so no canonical ordering of children is required. The decoder predicts part existence, semantic labels, leaf status, and sibling edges, then runs message passing along the predicted edges so parts refine and coordinate their geometry. The latent code $z \in \mathbb{R}^{256}$ of the root node represents the whole shape, and a structure consistency loss enforces that decoded geometries actually satisfy decoded symmetry and adjacency relationships, including relationships inherited by subtrees. The paper demonstrates that this joint representation supports interpolation that changes structure in small intuitive steps, part-preserving editing, and projection of unannotated images, point clouds, and partial scans into the same space to obtain structured abstractions.

Load-bearing premise

The paper assumes that the geometric relationships that matter for a plausible shape occur between sibling parts in the hierarchy, so symmetries or alignments linking parts in different subtrees can be ignored without losing realism.

Editorial extensions

If this is right

  • Generated shapes carry a full structured description—part hierarchy, semantics, and sibling relationships—so downstream users can edit, swap, or reuse individual parts directly.
  • Latent interpolation produces stepwise structural changes: a pedestal base shrinks before becoming four legs, a backrest loses one bar at a time, and each intermediate remains a functional chair.
  • Unannotated inputs such as images, synthetic point clouds, and partial real scans can be projected into the same latent space, recovering a structured shape with part segmentation and relationship edges.
  • A user can edit one part and search the latent space for the closest plausible shape satisfying the edit, so neighboring parts adjust to keep symmetries and adjacencies intact.
  • The approach scales to shape families with thousands of examples because it avoids the hard combinatorial search for a canonical binary tree that binary-recursive methods require.

Reading between the lines

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

  • A natural stress test is to add non-sibling relationship edges: the current decoding already predicts all $n_p^2$ sibling pairs, so extending to all pairs in a subtree would raise memory quadratically, but sparse or attention-based edge selection could make cross-subtree relationships affordable.
  • Because any modality-specific encoder can be trained to map into the fixed latent space, the same structure decoder could be reused for sketches, depth maps, or raw meshes, and the induced segmentations could be compared directly against annotated ground truth.
  • The sibling-only sparsification implies that a shape family whose critical constraints connect parts in different subtrees would expose the limitation; such families are the right test cases for whether the four edge types suffice.
  • Moving from objects to scenes—treating objects as parts and spatial relations as sibling edges, a direction the authors list as future work—would let the same graph machinery generate room layouts with plausible inter-object relationships.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 8 minor

Summary. The paper introduces StructureNet, a variational autoencoder for 3D shapes represented as hierarchies of n-ary graphs, where each node is a part or assembly and horizontal edges between siblings encode geometric relationships such as adjacency and symmetry. The encoder is recursive and order-invariant, using graph message passing along sibling edges and max-pooling to produce a 256-dimensional latent code; the decoder predicts a fixed maximum number of child parts and pairwise edge types, then applies message passing and geometry decoders for both bounding-box and point-cloud part representations. Training combines reconstruction losses (geometry, box normals, part/edge existence, semantics, leaf prediction), a structure-consistency loss that enforces symmetry and adjacency in the decoded shape, and a VAE regularizer. The paper reports experiments on PartNet chairs, tables, cabinets, and smaller categories, comparing reconstruction and generation against GRASS and a holistic PointNet++ autoencoder, and demonstrates applications in interpolation, abstraction from images/point clouds/partial scans, and structure-aware editing.

Significance. If the results hold, StructureNet makes a substantial contribution: it provides a single latent space that jointly captures continuous part geometry and discrete structural variation, without the binarization overhead of GRASS, and it explicitly models inter-part relationships during encoding and decoding. The architecture is clearly specified, the losses are well motivated, and the qualitative demonstrations are impressive—the stepwise structural interpolations in Figure 13 and the abstraction results in Figure 14 are particularly compelling. The paper also ships a large supplement with additional categories, semantic hierarchies, and extra results, which strengthens reproducibility. The main weaknesses are quantitative: the GRASS comparison is run on a reduced subset after discarding 840 chairs, the evaluation metrics (chamfer distance, consistency errors) overlap with StructureNet's own training losses, and Tables 1-3 contain point estimates without variance or significance tests. These issues do not invalidate the central idea but currently limit the strength of the claim of 'significant advantages over baseline'.

major comments (4)
  1. [Section 4, 'Geometric relationships'] The representation restricts geometric relationships to sibling pairs: the text states 'we choose to only capture geometric relationships between siblings in the hierarchy' and justifies it by saying non-sibling relationships are 'usually less significant or indirectly implied.' This assumption is load-bearing for the abstract claim that StructureNet generates 'realistic structured shape geometries,' because the graph encoder (Eq. 1), the edge predictor (Eq. 6), and both structure-consistency losses (Eqs. 20-22) operate exclusively on sibling edges and their descendant subtrees. The failure cases in Figure 16 (detached parts, asymmetric parts) are precisely the kinds of cross-subtree constraint violations that this representation cannot prevent. The paper provides no quantitative support for the 'less significant' claim. Please add an empirical analysis, e.g., the fraction of detected geometric relationships in PartNet that are between siblings versus non-siblings, and/or a direct metric of cross-subtree consistency (such as global reflection-symmetry error) on generated shapes. Without such evidence, the realism claim rests on an untested representational assumption.
  2. [Section 6.1, 'Comparison to GRASS' and Figure 7] The quantitative reconstruction comparison to GRASS is performed on a reduced subset of 4,031 chairs: the text notes that 'the GRASS authors reported that their pipeline failed to produce results for the remaining 840 chairs in our dataset.' The paper reports StructureNet's error on this reduced subset but does not report its error on the full 4,871-chair test set. If the excluded chairs are systematically harder (e.g., more complex structures or more unbalanced geometry), the comparison overstates StructureNet's advantage. Please report StructureNet's reconstruction errors separately on the full test set and on the common subset, and discuss the characteristics of the excluded chairs. This is necessary to make the 'significant improvement in reconstruction performance' claim verifiable.
  3. [Section 6.2, Eq. (25) and Table 2] The generation 'quality' and 'coverage' metrics are chamfer distances to the training set, and the reconstruction metrics EP, Erc, and Egc in Section 6.1 are non-squared analogues of StructureNet's own training losses Lgeo (Eq. 12) and Lsc (Eq. 22). Since neither GRASS nor the holistic autoencoder is trained with these objectives, the quantitative advantage of StructureNet in Tables 2 and Figure 7 may partly reflect alignment between the evaluation metric and StructureNet's training loss rather than superiority in perceptual or functional quality. To support the claim of 'significant advantages,' please add at least one independent evaluation—for example, a small user study of structural plausibility, or a metric based on global symmetry/part-contact satisfaction that is not directly a StructureNet training loss—or explicitly discuss this potential bias and its implications for the reported numbers.
  4. [Tables 1-3] The quantitative results are reported as point estimates without variance, number of repeated runs, or statistical significance tests. In Table 3, the differences are small (e.g., EP 0.0620 versus 0.0616, Erc 0.0183 versus 0.0186), and a reader cannot assess whether these differences are reliable or within run-to-run noise. Please report means and standard deviations over multiple training runs (or at least state the number of runs and the observed variability) for the main reconstruction, generation, and ablation tables. This is particularly important because the cross-method comparisons in Table 2 and Figure 7 involve absolute differences that are also small.
minor comments (8)
  1. [Eq. (25)] The equation for quality and coverage appears to have a formatting artifact: 'quality B ...' and 'coverage B ...' should read 'quality = ...' and 'coverage = ...'. Also, the set over which each average is taken (|SG| or |S|) should be written explicitly.
  2. [Section 5.3, loss list] The enumeration of loss components labels both 'Semantic loss' and 'Leaf loss' as '(v)'; the leaf loss should be '(vi)'.
  3. [Section 6.4, image and point cloud abstraction] The training objectives 'LI = e(S)−eI(render(S,θ))' and 'LO = e(S)−eO(sample(S))' are written as residuals without a norm or expectation. They should be expressed as distances to be minimized, e.g., LI = E[||e(S)−eI(render(S,θ))||^2].
  4. [Table 3] The 'no edges' row has incomplete entries: it omits ER (presumably undefined) and does not report consistency errors Erc/Egc. Please clarify whether these are intentionally left blank and how the reader should interpret the row.
  5. [Figure 7 caption] The caption contains a typo: 'GRASS Orginal' should be 'GRASS Original'.
  6. [Section 5.1, Eq. (1)] The message-passing update divides by M, the number of neighbors, but the case M=0 (a child part with no incident relationship edges) is not addressed; please state how isolated nodes are handled.
  7. [Section 6, data preparation] The definition of adjacency ('smallest distance is below 0.05∗r̄') and the quantity qmin in Eq. (21) are not fully specified for the point-cloud representation. Please clarify how qmin is computed for point clouds and how the threshold is applied when parts are represented as points.
  8. [Section 6.1] The sentence 'we train an non-variational autoencoder version' contains a grammar error ('an non-variational'); it should be 'a non-variational'.

Circularity Check

2 steps flagged · score 3.0 of 10

Evaluation metrics partly re-label the training losses, but the central generative architecture is not circularly derived.

  1. fitted input called prediction [Section 6.1, Table 1 (reconstruction and consistency errors), compared with Eq. 18 and Eq. 22]
    "The geometry reconstruction error EP is defined analogously to the geometry loss Lgeo, except that we use the non-squared chamfer distance. ... The reconstructed consistency error Erc measures the consistency of the reconstructed geometry with the reconstructed relationship edges. It is defined equivalent to Lsc, except that we use the non-squared chamfer distance."

    EP is Lgeo with the squared chamfer replaced by the non-squared chamfer, and Erc is Lsc up to the same substitution. Both Lgeo and Lsc are explicit terms in the training objective in Eq. 7, Eq. 18, and Eq. 22, so Table 1 reports, under the names of evaluation errors, how well the network minimized its own losses. Egc is partly independent because it uses the input relationship edges rather than the reconstructed ones, but EP and Erc are not external yardsticks. This is transparent and applied equally to all compared methods, and it does not by itself force the paper's main architectural conclusions.

  2. other [Section 6.2, Eq. 25 and Table 2, compared with Eq. 10 and Eq. 12]
    "The quality, of a generated shape set is measured by the average closest distance to any data sample, while the coverage is measured as the average closest distance from each data sample to a generated sample ... dS is the chamfer distance between the point representations of two shapes."

    The generation quality and coverage metrics use the same chamfer-distance family as the geometry reconstruction loss Lgeo (Eqs. 10 and 12) that trains the decoder to reproduce training shapes. Thus the quantitative generation scores partly reward the metric the model was trained to optimize. This is a metric-overlap concern rather than a forced prediction: the sampled latent codes are not directly optimized against Eq. 25, and the same distance is used to score GRASS, so the relative comparison remains meaningful. It is a mild self-referential element in the evaluation, not a derivation of the central result.

full rationale

StructureNet is an empirical architecture paper; there is no theorem or first-principles derivation chain whose conclusion can reduce to an input assumption. The shape representation, hierarchical graph encoder/decoder, and losses are introduced and then tested on held-out PartNet splits against GRASS and holistic baselines. The only concrete reduction I can exhibit is evaluative: EP and Erc are, up to chamfer squaring, the training losses Lgeo and Lsc, and the Eq. 25 quality/coverage metric uses the same chamfer distance as Lgeo. That makes the quantitative tables partly self-referential, but the overlap is transparent, applied symmetrically to baselines, and does not undermine the qualitative generation and interpolation evidence or the ablation study. The sibling-only relationship restriction in Section 4 is a substantive modeling assumption and a correctness risk for the realism claim, but it is not circular: it is a design choice stated before evaluation, not a conclusion derived from itself. PartNet is a self-citation by overlapping authors, but it is an externally published benchmark with fixed splits, so it is independent support rather than load-bearing circularity. Overall, the central claim has independent content; the circularity score reflects only the partial overlap between training objectives and evaluation metrics.

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

The central claim rests on the PartNet representation, the paper-specific sibling-only relationship assumption, and the fixed maximum arity. The losses introduce hand-chosen weights but no physically invented entities. The hyperparameters are standard for VAEs and graph networks.

free parameters (5)
  • beta (VAE regularizer weight) = 0.05
    Set empirically in Section 5.3 to balance reconstruction against a smooth latent space.
  • loss weights (alpha, gamma, lambda) = (20, 10, 0.1)
    Set empirically in Section 5.3, after Eq. 18, for geometry, normal, and semantic terms.
  • max decoded children np = 10
    Fixed maximum number of child parts decoded per node; also used as a dataset filter in Section 6.
  • adjacency threshold = 0.05 * mean bounding sphere radius
    Defines adjacency relationship edges from geometry in Section 6, Data Preparation.
  • message passing iterations = 2
    Two iterations used in both graph encoder and decoder, chosen experimentally (Sections 5.1 and 5.2).
assumptions (4)
  • domain assumption PartNet part hierarchies and semantic labels are correct ground truth for structure.
    The entire training signal (hierarchies, labels, geometry) is taken from PartNet (Mo et al. 2019), authored in part by this paper's authors; any errors propagate to the latent space.
  • ad hoc to paper Important geometric relationships occur between siblings; non-sibling relations can be neglected.
    Stated in Section 4: 'we choose to only capture geometric relationships between siblings in the hierarchy.' This sparsifies the graph but may drop cross-subtree constraints.
  • ad hoc to paper A fixed maximum of 10 children per node is sufficient to represent the dataset after filtering.
    Shapes with more than 10 children in any part are removed in Section 6; the method cannot represent them, a stated limitation.
  • domain assumption Symmetry relationships provided by the method of Wang et al. 2011 are accurate.
    Relationship edges are precomputed using external symmetry detection in Section 6; missed or false symmetries become incorrect supervision for edge prediction and Lsc.

how reviews work

0 comments
Cite this review

Pith. "Pith review of StructureNet: Hierarchical Graph Networks for 3D Shape Generation." pith.science (2026). https://pith.science/paper/IELKIP5D

@misc{pith2026190800575,
  author       = {Pith},
  title        = {Pith review of: StructureNet: Hierarchical Graph Networks for 3D Shape Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IELKIP5D}},
  note         = {Machine review of arXiv:1908.00575}
}
read the original abstract

The ability to generate novel, diverse, and realistic 3D shapes along with associated part semantics and structure is central to many applications requiring high-quality 3D assets or large volumes of realistic training data. A key challenge towards this goal is how to accommodate diverse shape variations, including both continuous deformations of parts as well as structural or discrete alterations which add to, remove from, or modify the shape constituents and compositional structure. Such object structure can typically be organized into a hierarchy of constituent object parts and relationships, represented as a hierarchy of n-ary graphs. We introduce StructureNet, a hierarchical graph network which (i) can directly encode shapes represented as such n-ary graphs; (ii) can be robustly trained on large and complex shape families; and (iii) can be used to generate a great diversity of realistic structured shape geometries. Technically, we accomplish this by drawing inspiration from recent advances in graph neural networks to propose an order-invariant encoding of n-ary graphs, considering jointly both part geometry and inter-part relations during network training. We extensively evaluate the quality of the learned latent spaces for various shape families and show significant advantages over baseline and competing methods. The learned latent spaces enable several structure-aware geometry processing applications, including shape generation and interpolation, shape editing, or shape structure discovery directly from un-annotated images, point clouds, or partial scans.

Figures

Figures reproduced from arXiv: 1908.00575 by the authors.

Figure 1
Figure 1. StructureNet is a hierarchical graph network that produces a unified latent space to encode structured models with both continuous geometric and discrete structural variations. In this example, we projected an un-annotated point cloud (left) and un-annotated image (right) into the learned latent space yielding semantically segmented point clouds structured as a hierarchy of graphs. The shape interpolation in the lat… view at source ↗
Figure 2
Figure 2. N-ary part hierarchies. Shape parts can naturally be organized into n-ary hierarchies. Here we show the part hierarchies of two shapes as defined by PartNet [Mo et al. 2019]. The top row shows oriented bounding boxes of leaf parts and the hierarchy is illustrated below. Hierarchy nodes have the same color as the corresponding part. Note how geometrically dissimilar shapes may have consistent hierarchies. Our shape r… view at source ↗
Figure 3
Figure 3. Interpolation compared to GRASS. We compare interpolations between several pairs of chairs using StructureNet (colored boxes), and using GRASS (blue boxes). We interpolate between shapes from our test set shown on the left-most and right-most sides, after being reconstructed by both methods (marked as ‘source’ and ‘target’). Our interpolations use a larger number of smaller structural changes to reach the target sha… view at source ↗
Figures from the paper (23 more)
Figure 4
Figure 4. Figure 4: Shape Representation. Shapes are represented by their hierarchi￾cal decomposition into parts (black edges), with geometric relationships between siblings (orange arrows): adjacency (τa), translational symmetry (τt ), reflective symmetry (τr ), and rotational symmetry (…
Figure 5
Figure 5. Figure 5: Hierarchical Graph Networks. Our variational autoencoder consists of two encoders and two decoders that both operate on our shape representation. The geometry encoder egeo encodes the geometry of a part into a fixed-length feature vector f , illustrated with a gray cir…
Figure 6
Figure 6. Figure 6: Relationships between subtrees. A relationship between two non￾leaf parts also holds for their subtrees. The reflective symmetry τr of the parent parts on the left also holds for their children on the right. two subtrees to be mirrored in the same way, see [PITH_FULL_…
Figure 8
Figure 8. Figure 8: Reconstruction compared to a holistic approach. We recon￾struct the point cloud representation of the shapes in the top row with a holistic approach (orange) that does not use structure, and compare to StructureNet. The holistic approach suffers from noise that results…
Figure 9
Figure 9. Figure 9: Novelty of generated shapes. In the first column we show three generated shapes, and on the right the five closest matches in the training set, measured with the chamfer distance. Our generated structure and geometry is different from the shapes in the training set. at…
Figure 13
Figure 13. Figure 13: The left half of the figure shows interpolations between [PITH_FULL_IMAGE:figures/full_fig_p011_13.png]
Figure 10
Figure 10. Figure 10: Generated shapes. We show shapes in all categories decoded from random latent vectors, including shapes with bounding box geometry, and shapes with point cloud geometry. Parts are colored according to semantics, see the Supplementary for the full semantic hierarchy fo…
Figure 12
Figure 12. Figure 12: Part Interpolation. We interpolate either only the backrest (first row) or only the base (second row) between the chairs on the left and right side. Intermediate shapes preserve structural plausibility of the interpolated result mainly through geometric differences to…
Figure 13
Figure 13. Figure 13: Interpolation of shapes. We show interpolations between a source shape and a target shape from all categories. Interpolations are symmetric, so source and target are interchangeable. Interpolations between shapes with box geometry is shown on the left side, and point …
Figure 15
Figure 15. Figure 15: Structure-aware part editing. We show editing results on two shapes with box geometry (first four rows) and two shapes with point cloud geometry (two bottom rows). For the two shapes with box geometry, we perform five different edits each, one edit per column. The edi…
Figure 16
Figure 16. Figure 16: Failure cases analysis. We present several failure cases we ob￾served for box-shape and point cloud generation. We see discrete errors such as missing parts (e.g. first row, first column), duplicate parts (e.g.second row, second column), detached parts (e.g. first row…
Figure 17
Figure 17. Figure 17: Image and point cloud retrieval. Images and point clouds are retrieved using on a query shape based on the distance from the query in the multi-modal latent space. The bottom three rows compare retrieval with the full shape as query to part-based retrieval using the b…
Figure 18
Figure 18. Figure 18: Joint embedding of images, point clouds and shapes. We visualize the multi-modal latent space as a two-dimensional embedding. At each grid point, we randomly show one of the modalities [PITH_FULL_IMAGE:figures/full_fig_p020_18.png]
Figure 19
Figure 19. Figure 19: PartNet semantic hierarchy for chairs. Dash lines show the OR-nodes and solid lines show the AND-node in PartNet. We assign the semantic labels in the figures with the colors that we use for box-shape and point cloud visualization in the main paper [PITH_FULL_IMAGE:f…
Figure 20
Figure 20. Figure 20: PartNet semantic hierarchy for tables. Dash lines show the OR-nodes and solid lines show the AND-node in PartNet. We assign the semantic labels in the figures with the colors that we use for box-shape and point cloud visualization in the main paper [PITH_FULL_IMAGE:f…
Figure 21
Figure 21. Figure 21: PartNet semantic hierarchy for storage furnitures. Dash lines show the OR-nodes and solid lines show the AND-node in PartNet. We assign the semantic labels in the figures with the colors that we use for box-shape and point cloud visualization in the main paper [PITH_…
Figure 22
Figure 22. Figure 22: Shape generation results for vases and trash cans. The datasets for these categories are smaller than for our main categories: 505 samples for vases and 83 for trashcans. Vases have a less complex structure compared to the other categories, making the quality of the g…
Figure 23
Figure 23. Figure 23: Shape interpolation results for vases and trash cans. Similar to our main categories, structure is interpolated smoothly. The last rows for vases and trash cans show that the part geometry is interpolated smoothly as well [PITH_FULL_IMAGE:figures/full_fig_p024_23.png]
Figure 24
Figure 24. Figure 24: Shape generation results for beds. We also test on this third, severely under-sampled category, with a training set size of 54. As we can see in [PITH_FULL_IMAGE:figures/full_fig_p025_24.png]
Figure 25
Figure 25. Figure 25: More Box-shape Generation Results [PITH_FULL_IMAGE:figures/full_fig_p026_25.png]
Figure 26
Figure 26. Figure 26: More Point Cloud Generation Results [PITH_FULL_IMAGE:figures/full_fig_p027_26.png]
Figure 27
Figure 27. Figure 27: More Shape Interpolation Results [PITH_FULL_IMAGE:figures/full_fig_p028_27.png]
Figure 28
Figure 28. Figure 28: More Shape Abstraction Results [PITH_FULL_IMAGE:figures/full_fig_p029_28.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 7 Pith papers

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

  1. Articulate AnyMesh: Open-Vocabulary 3D Articulated Objects Modeling

    cs.CV 2025-02 conditional novelty 7.0 of 10

    Articulate AnyMesh converts arbitrary rigid 3D meshes into articulated objects by combining VLM-driven part segmentation, geometry-aware joint estimation, and optional shape completion.

  2. SAR2Struct: Extracting 3D Semantic Structural Representation of Aircraft Targets from Single-View SAR Image

    cs.CV 2025-06 conditional novelty 6.0 of 10

    SAR2Struct recovers a 3D hierarchical part structure of aircraft from a single radar image by detecting 2D component keypoints and mapping them to a symmetry hierarchy with a graph neural network and recursive decoder.

  3. PartCrafter: Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers

    cs.CV 2025-06 conditional novelty 6.0 of 10

    PartCrafter generates several separable 3D part meshes at once from a single image by fine-tuning a pretrained 3D diffusion transformer with part identity tokens and local-global attention.

  4. MeshArt: Generating Articulated Meshes with Structure-Guided Transformers

    cs.CV 2024-12 conditional novelty 6.0 of 10

    A hierarchical transformer that tokenizes triangles generates articulated 3D meshes part-by-part, with 57.1% higher structure coverage and a 209-point lower mesh FID than prior methods on an expanded PartNet.

  5. Wavelet Latent Diffusion (Wala): Billion-Parameter 3D Generative Model with Compact Wavelet Encodings

    cs.CV 2024-11 conditional novelty 6.0 of 10

    Wavelet Latent Diffusion (WaLa) shrinks 3D shapes to 6,912-variable latent codes and trains billion-parameter diffusion models that generate 256^3 geometry in 2-4 seconds, claiming state-of-the-art results.

  6. CAD-Coder:Text-Guided CAD Files Code Generation

    cs.GR 2025-05 conditional novelty 5.0 of 10

    A fine-tuned LLM converts template-style text prompts into ezdxf Python code that produces editable, annotated Dxf CAD files.

  7. Efficient Sensorimotor Learning for Open-world Robot Manipulation

    cs.RO 2025-05 conditional novelty 4.0 of 10

    A PhD dissertation argues that object, spatial, and behavioral regularities, extracted with foundation models, enable data-efficient, generalizable robot manipulation, and presents seven systems and a benchmark built ...

Reference graph

Works this paper leans on

8 extracted references · 5 canonical work pages · cited by 7 Pith papers

  1. [5]

    ACM Transactions on Graphics 36, 4 (2017)

    GRASS: Generative Recursive Autoencoders for Shape Structures. ACM Transactions on Graphics 36, 4 (2017). Manyi Li, Akshay Gadi Patil, Kai Xu, Siddhartha Chaudhuri, Owais Khan, Ariel Shamir, Changhe Tu, Baoquan Chen, Daniel Cohen-Or, and Hao Zhang. 2019a. Grains: Generative recursive autoencoders for indoor scenes. ACM Transactions on Graphics (TOG) 38, 2...

    arXiv 2017

  2. [7]

    ACM Transactions on Graphics (TOG) 36, 6 (2017), 226

    Complementme: weakly-supervised component suggestions for 3D modeling. ACM Transactions on Graphics (TOG) 36, 6 (2017), 226. Maxim Tatarchenko, Alexey Dosovitskiy, and Thomas Brox. 2017. Octree generating networks: Efficient convolutional architectures for high-resolution 3d outputs. In Proceedings of the IEEE International Conference on Computer Vision ....

    arXiv 2017

  3. [8]

    Learning Hierarchical Shape Segmentation and Labeling from Online Repositories

    SAGNet: Structure-aware Generative Network for 3D-Shape Modeling. ACM Transactions on Graphics (Proceedings of SIGGRAPH 2019) 38, 4 (2019), 91:1–91:14. Zhige Xie, Kai Xu, Ligang Liu, and Yueshan Xiong. 2014. 3d shape segmentation and labeling via extreme learning machine. In Computer graphics forum, Vol. 33. Wiley Online Library, 85–95. Keyulu Xu, Weihua ...

    work page Pith review arXiv 2019

  4. [2011]

    In ACM Transactions on Graphics (TOG), Vol

    Probabilistic reasoning for assembly-based 3D modeling. In ACM Transactions on Graphics (TOG), Vol. 30. ACM, 35. Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 2016. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In European conference on computer vision . Springer, 628–644. Angela Dai, Ange...

    work page 2016

  5. [2012]

    InAdvances in neural information processing systems

    Convolutional-recursive deep learning for 3d object classification. InAdvances in neural information processing systems . 656–664. Richard Socher, Cliff C Lin, Chris Manning, and Andrew Y Ng. 2011. Parsing natural scenes and natural language with recursive neural networks. In Proceedings of the 28th international conference on machine learning (ICML-11) ....

    work page 2011

  6. [2017]

    InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    3D shape segmentation with projective convolutional networks. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition . 3779–3788. Evangelos Kalogerakis, Siddhartha Chaudhuri, Daphne Koller, and Vladlen Koltun. 2012. A probabilistic model for component-based shape synthesis. ACM Transactions on Graphics (TOG) 31, 4 (2012), 55. Evan...

    work page 2012

  7. [2018]

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

    A papier-mâché approach to learning 3d surface generation. In Proceedings of the IEEE conference on computer vision and pattern recognition . 216–224. JunYoung Gwak, Christopher B Choy, Manmohan Chandraker, Animesh Garg, and Silvio Savarese. 2017. Weakly Supervised 3D Reconstruction with Adversarial Constraint. In 3D Vision (3DV), 2017 Fifth International...

    work page 2015

  8. [2019]

    Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2019)

    ShapeGlot: Learning Language for Shape Differentiation. Proceedings of the IEEE International Conference on Computer Vision (ICCV) (2019). Amir Arsalan Soltani, Haibin Huang, Jiajun Wu, Tejas D Kulkarni, and Joshua B Tenen- baum. 2017. Synthesizing 3d shapes via modeling multi-view depth maps and silhouettes with deep generative networks. In Proceedings o...

    arXiv 2019

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.