Pith. sign in

REVIEW 4 major objections 4 minor 57 references

SARe: Structure-Aware Generative 3D Fragment Reassembly

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read SARe claims that the bottleneck in 3D fragment reassembly at large fragment counts is contact-structure uncertainty, and that jointly predicting fracture-surface labels and an adjacency graph during flow-based assembly, then verifying and r

desk verdict Solid and useful paper on generative fragment reassembly; the structural supervision is clearly helpful, but the inference-time refinement gains are modest and not significance-tested, so the 'actionable' contact graph claim is only partially supported. read the letter →

arxiv 2603.21611 v2 pith:NHIN6GLA submitted 2026-03-23 cs.CV

Hanze Jia , Chunshi Wang , Yuxiao Yang , Zhonghua Jiang , Yawei Luo , Shuainan Ye , Tan Tang This is my paper
classification cs.CV
keywords 3Dfragmentreassemblyrectifiedflowcontactgraphpredictionfracturesurfacesegmentationinference-timerefinementgenerativepointcloudassemblypartaccuracyposeestimation
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

This paper tries to establish that the reason 3D fragment reassembly fails at large fragment counts is not pose estimation alone but unreliable contact structure: which fracture surfaces meet and which fragments are adjacent. The authors show that as fragment count grows, adjacency recall of predicted assemblies drops, and that an oracle second pass conditioned on a few ground-truth adjacencies improves results—evidence that contact errors cascade. Their method, SARe, makes contact structure an explicit predicted object rather than a byproduct: a rectified-flow generator classifies fracture-surface tokens and an inter-fragment adjacency graph while transporting query points into a shared frame, and an inference-time refinement stage verifies candidate edges geometrically, freezes reliable substructures, and resamples uncertain fragments. Across synthetic, scanned-simulated, and physically fractured objects with 2 to 50 fragments, SARe reports higher part accuracy and slower degradation as fragment count increases, with the largest gains in the many-fragment regime. If right, the paper's recipe—joint structural supervision plus geometry-verified inference-time correction—is a general path to scalable assembly.

What carries the argument

Query-point-aligned local geometric tokens—extracted from a frozen pretrained shape encoder at farthest-point-sampled locations, augmented with coordinates, normals, and a fragment identity embedding—serve as conditioning for a rectified-flow velocity field parameterized by a transformer operating on assembled query point tokens. Two lightweight structural heads attached to an intermediate layer predict per-token fracture-surface probabilities and a symmetric contact-adjacency graph, trained with binary cross-entropy alongside the flow-matching objective. At inference, SARe-Refine verifies candidate edges by voxelizing predicted poses, rejecting pairs whose overlap ratio exceeds a threshold,

What would settle it

Replace the frozen pretrained encoder's local tokens with raw farthest-point-sampled coordinates and surface normals, keeping the generator and training procedure otherwise identical, and compare part accuracy at the largest fragment count (K=50). If accuracy holds, the pretrained tokens are not the load-bearing component; if it collapses, the paper's dependence on those tokens is confirmed. A complementary check: measure adjacency precision on thin, slice-like fragments, where the authors report the method struggles.

Watch

Extended reading notes

Core claim

The paper claims that the scalability bottleneck in 3D fragment reassembly is not pose estimation alone but uncertainty about which fragments touch: adjacency recall drops as fragment count grows, and an oracle experiment shows that conditioning a second inference on a handful of ground-truth adjacencies improves the assembly. SARe therefore makes contact structure an explicit, predicted object: SARe-Gen, a rectified-flow generator that moves query points on fragment surfaces into a shared frame, attaches two lightweight heads to an intermediate transformer representation, one classifying each query token as fracture surface and one scoring fragment-pair adjacency, and trains them jointly wi

Load-bearing premise

The approach assumes that the frozen pretrained encoder's local surface tokens, which were never trained on broken edges, still carry enough detail about fracture surfaces for the flow model and structural heads to work with; if those tokens blur away high-frequency fracture features, the whole large-fragment advantage disappears.

Editorial extensions

If this is right

  • Explicit adjacency prediction turns contact structure into an actionable constraint, so assembly errors can be corrected at inference time instead of being accepted as a byproduct of pose regression.
  • Because structural heads are trained jointly with the flow objective on intermediate tokens, the method needs no separate verifier network or teacher-alignment stage.
  • As fragment count grows from 2 to 50, SARe's part accuracy declines more slowly than pose-only flow and diffusion baselines, with the gap widening in high-K bins.
  • On the hardest test subset (first-pass accuracy at or below 95 percent), SARe-Refine adds roughly 2.5–3 percentage points in part accuracy across three data settings, while barely changing already-accurate assemblies.
  • Replacing the frozen shape encoder with a point encoder drops part accuracy substantially in the ablation, indicating the local token substrate is a necessary component of the pipeline.

Reading between the lines

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

  • Extrapolating from the oracle diagnostic, the paper's logic implies that the remaining headroom at large K is bounded by contact-graph recall: a perfect adjacency predictor would still leave pose-level errors, but the cascade from misidentified adjacencies would be eliminated. This could be tested by feeding ground-truth edges into SARe-Refine at K=50 and measuring the residual performance gap.
  • The same query-aligned token plus structural-supervision recipe could transfer to semantic part assembly or 2D/3D puzzle solving, where explicit adjacency is also a natural target; the paper does not test those settings.
  • A hidden dependence is the frozen encoder's sensitivity to fracture geometry; a cheap probe would be to evaluate adjacency precision on thin, slice-like fragments, where the paper reports failures, and to compare token reconstructions of fracture surfaces against intact surfaces.
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 introduces SARe, a generative framework for 3D fragment reassembly in which a rectified-flow generator (SARe-Gen) jointly predicts assembled query points, token-level fracture-surface probabilities, and an inter-fragment contact graph, followed by an optional inference-time refinement stage (SARe-Refine) that verifies predicted adjacencies, identifies reliable substructures, and resamples uncertain fragments in a RePaint-style second pass. The method is evaluated on Breaking Bad, Fantastic Breaks, and a new OmniObject3D-derived fracture benchmark, reporting state-of-the-art part accuracy, Chamfer distance, and pose errors, with ablations supporting the value of the two structural heads and the query-aligned ShapeVAE tokens. The paper argues that explicit contact-structure modeling is the key to scalability in the many-fragment regime and that the predicted graph is actionable at inference rather than a byproduct.

Significance. If the results hold, the paper makes a useful contribution: it provides evidence that auxiliary structural supervision on fracture surfaces and adjacency can improve generative point-flow assembly, and it proposes a concrete inference-time mechanism that uses predicted structure to correct uncertain fragments. The evaluation is unusually broad — three datasets spanning synthetic, simulated-scan, and real-scanned fractures — and the ablations (Tables 3–5) are internally consistent. The oracle diagnostic in Fig. 1(c) is appropriately labeled as a diagnostic and is not used as an evaluation target, so there is no circularity in the main claims. However, the central claim that the contact graph is actionable at inference rests on modest refinement gains that currently lack statistical support, and several load-bearing details are deferred to a supplementary appendix that is not present in the submitted manuscript. The paper also does not ship code or data release links, which limits reproducibility despite the stated intent to release a benchmark.

major comments (4)
  1. [§5.2, Table 2 and Table 5] The evidence that SARe-Refine is 'actionable' is statistically under-powered. The full-test gains are +0.62 pp (Everyday), +0.65 pp (Artifact), and +1.01 pp (OmniObject3D); hard-subset gains are +2.49–3.06 pp. These are reported without error bars, number of seeds, or significance tests. Table 5 shows that the RePaint resampling mode gains +0.62 pp over no refinement while simply freezing the verified substructure (Freeze) gains +0.11 pp; the additional benefit attributable to resampling is therefore about 0.5 pp, which could be within seed/initialization variance. Since the abstract and contribution (2) claim that the predicted graph guides inference-time correction, please add multi-seed runs with confidence intervals, paired significance tests, and a per-bin breakdown for SARe-Refine.
  2. [§5.2, Fig. 1 and Table 1] The central claim of 'more graceful degradation in challenging many-fragment settings' is not backed by the quantitative presentation. Fig. 1(a) is referenced but no numeric K-stratified PA or CD values are given in the text; Fig. 1(b) uses an induced adjacency matrix whose threshold is deferred to the missing supplementary; and no K-stratified results are reported for SARe-Refine. Because the large-K advantage is the paper's main differentiator, please provide a table or figure with PA/CD and adjacency recall per fragment-count bin, with error bars, and report the K distribution of the test sets.
  3. [§4.1, Eq. (2), and Table 3] The entire conditioning path rests on the frozen ShapeVAE query-aligned tokens, which were not trained on fracture surfaces. The only evidence for this substrate is the ShapeVAE→PTv3 swap in Table 3, but that swap also changes the tokenization/conditioning protocol (RPF-style feature sampling), so it does not isolate token fidelity. The paper should directly validate that the local tokens preserve fine-grained fracture-surface geometry: report fracture-token precision/recall on the validation set, or compare frozen encoders under the exact same query-aligned extraction scheme. Without this, the proposed structural heads may be latching onto a representation that does not generalize to unseen fracture patterns.
  4. [§5.1, §4.3, and 'supplementary material' references] Several load-bearing implementation details are promised in the supplementary material but are absent from the submitted manuscript: dataset generation parameters for OmniObject3D, the induced-adjacency threshold used for baselines in Fig. 1(b), the geometric-verification thresholds in SARe-Refine (τ_o, voxel tolerance, minimum component size), the per-fragment query-point allocation M_i, and the resampling schedule. Parameters such as τ_o and α are free choices that could materially affect the refinement result. Please include these details in the manuscript or a supplementary appendix, and provide the missing benchmark release link.
minor comments (4)
  1. [Table 1] Header typo: 'F antastic Breaks' should be 'Fantastic Breaks'. Also, the notation SARe-Gen∗ vs SARe-Gen vs SARe is confusing; the distinction between Breaking-Bad-only training, joint training, and the refinement stage should be defined in the table caption or Sec. 5.1.
  2. [Table 2] 'P ASARe ≤95%' is awkward; use 'first-pass PA ≤ 95%' consistently.
  3. [Fig. 4] The figure contains 'SGRe-Refine' as an apparent typo for 'SARe-Refine'. Also, failure cases are mentioned but the figure legend does not clearly mark which panels are failures; please make this explicit.
  4. [§5.4] The statement in Table 4 discussion that deeper features are 'more specialized for pose generation' is an interpretation; the table only shows PA and adjacency precision, not a direct measure of feature specialization. Soften the wording or provide evidence.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SARe's structural predictions are supervised on ground-truth labels and evaluated on held-out data; no derivation step reduces to its inputs.

full rationale

This is an empirical pipeline rather than a derivation, and I find no step where a claimed prediction is equivalent to a fitted input by construction. The core objective in Eq. (4) is standard conditional rectified flow, and Eq. (5) adds structural heads trained with ground-truth supervision: the paper states that "Ground-truth labels for the fracture-surface token and the contact graph are obtained from dataset preprocessing." Evaluation metrics (PA, CD, RMSE) are computed by comparing predicted poses to ground-truth poses, independently of the predicted adjacency or fracture heads, so the structural predictions are not used to define the evaluation target. SARe-Refine uses predicted candidate edges filtered by geometric verification, not ground-truth adjacencies, and Eq. (6) is a RePaint-style blending operation; its modest reported gains and lack of significance tests are an empirical-support concern, not circularity. The Fig. 1-c experiment is explicitly labeled an "oracle diagnostic" and is used only as motivation, not presented as a predictive result. The frozen ShapeVAE conditioning is an external pretrained encoder and is not justified by a self-citation chain; even if it lacks fine-grained fracture detail, that is a representation-quality assumption, not a circular one. I found no load-bearing self-citations, no uniqueness theorem imported from the authors, and no ansatz smuggled in exclusively via the authors' own prior work. The skeptic attack about the contact graph being merely a regularizer is a causal-attribution issue about ablations, not a reduction of the central claim to its inputs.

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

The paper introduces no new physical entities. Its contribution is an architecture with ordinary ML hyperparameters and a verification threshold. The more consequential assumptions are the fitness of the frozen encoder and the validity of the induced adjacency baseline used to motivate the method.

free parameters (5)
  • Auxiliary loss weights λ_F, λ_A = 0.01 each
    Chosen by hand for the fracture and adjacency heads; no sweep or sensitivity analysis reported in the main text.
  • RePaint blending strength α = 0.5
    Fixed hyperparameter for SARe-Refine; Table 5 shows α=0.5 gives slightly higher PA than α=1.0.
  • Voxel overlap threshold τ_o = Not reported in main text
    Used in geometric verification to reject interpenetrating fragment pairs; value deferred to the appendix.
  • Total query budget M = 5120 per object
    Per-object query-point budget, allocated by surface area; affects model capacity and runtime.
  • Structural head attachment layer ℓ_s = 4
    Chosen based on Table 4 as a trade-off between PA and adjacency precision.
assumptions (5)
  • domain assumption Frozen ShapeVAE local tokens capture fracture-surface geometry
    Section 4.1 uses Hunyuan3D-ShapeVAE as a fixed encoder; the paper does not validate fracture-specific fidelity of these tokens.
  • standard math Rectified-flow linear interpolation with Gaussian endpoint is a suitable generative model for assembled point sets
    Equations (3)-(4) follow RPF; standard flow-matching, but still a modeling choice for this task.
  • domain assumption Auxiliary structural losses at an intermediate DiT layer shape the representations that drive coordinate transport
    Core training premise of the paper; supported only by ablations, not by representation analysis.
  • domain assumption The induced adjacency graph from baseline assemblies is a fair proxy for contact-structure quality
    Used in Fig. 1b to claim adjacency recall drops; thresholding minimum surface distances is not the same as the true contact graph.
  • domain assumption Fracture simulation for OmniObject3D produces fracture surfaces comparable to Breaking Bad
    Section 5.1 introduces a new benchmark using 'the same pipeline as Breaking Bad [37]'; the realism of the simulated fractures is not independently verified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SARe: Structure-Aware Generative 3D Fragment Reassembly." pith.science (2026). https://pith.science/paper/NHIN6GLA

@misc{pith2026260321611,
  author       = {Pith},
  title        = {Pith review of: SARe: Structure-Aware Generative 3D Fragment Reassembly},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NHIN6GLA}},
  note         = {Machine review of arXiv:2603.21611}
}
read the original abstract

3D fragment reassembly estimates the rigid pose of each fragment to recover a complete object from unordered point clouds or meshes. The task becomes increasingly challenging as the fragment count grows, since irregular fragments provide weak semantic cues and admit rapidly increasing numbers of plausible contact relations and global configurations. We propose Structure-Aware Reassembly (SARe), a generative framework that integrates query-aligned local geometry and task-native structural supervision into point-flow assembly. SARe-Gen conditions each transported surface query on a local latent and jointly supervises intermediate flow tokens with query-level fracture-region and fragment-level contact targets. Because these heads are optimized together with flow matching, structural supervision directly shapes the representations that drive coordinate transport, without requiring additional reassembly-specific pretraining or a separate teacher-alignment stage. At inference time, SARe-Refine geometrically verifies predicted relations and uses reliable local subassemblies to guide a second sampling pass, reinforcing consistent regions while resampling uncertain fragments. We evaluate SARe across three settings, including synthetic fractures, simulated fractures from scanned real objects, and scans of physically fractured objects. The results demonstrate state-of-the-art performance, with higher part accuracy and more graceful degradation in challenging many-fragment settings.

Figures

Figures reproduced from arXiv: 2603.21611 by the authors.

Figure 1
Figure 1. As the fragment count K increases, (a) part accuracy drops and (b) adja￾cency recall of the induced contact graph decreases accordingly, indicating growing contact-structure errors at larger scales. (c) As an oracle diagnostic, re-inference of RPF conditioned on a small set of GT adjacencies improves an example assembly. All results are computed on the Breaking Bad dataset. 14, 38, 57]. Given a set of fragment point… view at source ↗
Figure 2
Figure 2. Overview of SARe. SARe-Gen generates assembled query points in Euclidean space while predicting fracture-surface tokens and a contact graph. SARe-Refine op￾tionally refines hard cases by keeping reliable substructures fixed and resampling un￾certain regions. To address these limitations, we reformulate the assembly problem at two com￾plementary levels. Inspired by [16, 39, 55], we formulate reassembly as conditional… view at source ↗
Figure 3
Figure 3. SARe-Refine pipeline. Filter predicted edges, keep reliable substructures, then RePaint-style resample uncertain regions conditioned on a stable mask. transformer token representation h (ℓs) ∈ RM×D, where ℓs is the attachment layer index. Following representation supervision [51] and auxiliary multi-task diffu￾sion [48–50], these auxiliary structural targets encourage noise-robust, structure￾aware features. We predi… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative results of SARe-Gen and SARe-Refine, along with representative failure cases. (Please zoom in for a clearer view of the geometric details.) Removing the adjacency head yields a larger drop in PA, highlighting the impor￾tance of explicit adjacency prediction…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 6 canonical work pages

  1. [1]

    Calzado-Martínez, A., García-Fernández, A.L., Ortega-Alvarado, L.M., Feito- Higueruela, F.R.: Integrated information system for 3d interactive reconstruc- tion of an archaeological site. J. Comput. Cult. Herit.16(3) (2023).https: //doi.org/10.1145/3586077

    work page doi:10.1145/3586077 2023

  2. [2]

    In: Proceedings of the 12th International Conference on Virtual Reality, Archaeology and Cultural Heritage

    Castañeda, A.G., Brown, B., Rusinkiewicz, S., Funkhouser, T., Weyrich, T.: Global consistency in the automatic assembly of fragmented artefacts. In: Proceedings of the 12th International Conference on Virtual Reality, Archaeology and Cultural Heritage. p. 73–80. Eurographics Association (2011)

    2011

  3. [3]

    In: NeurIPS (2025),https://openreview.net/forum?id=ljJGBcpn7q

    Chen, M., Wang, J., Shapovalov, R., Monnier, T., Jung, H., Wang, D., Ranjan, R., Laina, I., Vedaldi, A.: Autopartgen: Autoregressive 3d part generation and discovery. In: NeurIPS (2025),https://openreview.net/forum?id=ljJGBcpn7q

    2025

  4. [4]

    In: CVPR

    Chen, R., Zhang, J., Liang, Y., Luo, G., Li, W., Liu, J., Li, X., Long, X., Feng, J., Tan, P.: Dora: Sampling and benchmarking for 3d shape variational auto-encoders. In: CVPR. pp. 16251–16261 (June 2025)

    2025

  5. [5]

    In: CVPR

    Chen, Y.C., Li, H., Turpin, D., Jacobson, A., Garg, A.: Neural shape mating: Self- supervised object assembly with adversarial shape priors. In: CVPR. pp. 12724– 12733 (June 2022)

    2022

  6. [6]

    ACMJournalonComputingandCulturalHeritage4(2)(2011).https://doi.org/ 10.1145/2037820.2037824

    Funkhouser, T., Shin, H., Toler-Franklin, C., García Castañeda, A., Brown, B., Dobkin,D.,Rusinkiewicz,S.,Weyrich,T.:Learninghowtomatchfrescofragments. ACMJournalonComputingandCulturalHeritage4(2)(2011).https://doi.org/ 10.1145/2037820.2037824

    arXiv 2011

  7. [7]

    In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., Sabato, S

    Ghasemipour, S.K.S., Kataoka, S., David, B., Freeman, D., Gu, S.S., Mordatch, I.: Blocks assemble! learning to assemble with large-scale structured reinforcement learning. In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., Sabato, S. (eds.) Proceedings of the 39th International Conference on Machine Learning. Proceedings of Machine Learnin...

    2022

  8. [8]

    In: CVPR

    Go, H., Park, B., Jang, J., Kim, J.Y., Kwon, S., Kim, C.: Splatflow: Multi-view rectified flow model for 3d gaussian splatting synthesis. In: CVPR. pp. 21524–21536 (June 2025)

    2025

Show all 57 references
  1. [9]

    In: CVPR

    Guo, L., Wang, C., Yang, W., Huang, S., Wang, Y., Pfister, H., Wen, B.: Shad- owdiffusion: When degradation prior meets diffusion model for shadow removal. In: CVPR. pp. 14049–14058 (2023)

    2023

  2. [10]

    In: ICCV

    Hong, J.H., Yoo, S.J., Zeeshan, M.A., Kim, Y.M., Kim, J.: Structure-from-sherds: Incremental 3d reassembly of axially symmetric pots from unordered and mixed fragment collections. In: ICCV. pp. 5423–5431 (2021).https://doi.org/10.1109/ ICCV48922.2021.00539

    2021

  3. [11]

    In: NeurIPS (2020)

    Huang, J., Zhan, G., Fan, Q., Mo, K., Shao, L., Chen, B., Guibas, L., Dong, H.: Generative 3d part assembly via dynamic graph learning. In: NeurIPS (2020)

    2020

  4. [12]

    ACM TOG25(3), 569–578 (2006)

    Huang, Q.X., Flöry, S., Gelfand, N., Hofer, M., Pottmann, H.: Reassembling fractured objects by geometric matching. ACM TOG25(3), 569–578 (2006). https://doi.org/10.1145/1141911.1141925

    2006 arXiv

  5. [13]

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

    2025 arXiv

  6. [14]

    Sensors25(21) (2025)

    Jia, C., Ren, Y., Wang, Z., Zhang, Y.: Geoassemble: A geometry-aware hierarchical method for point cloud-based multi-fragment assembly. Sensors25(21) (2025). https://doi.org/10.3390/s25216533,https://www.mdpi.com/1424-8220/25/ 21/6533 16 H. Jia et al

    2025 doi

  7. [15]

    In: CVPR

    Lamb, N., Palmer, C., Molloy, B., Banerjee, S., Banerjee, N.K.: Fantastic breaks: A dataset of paired 3d scans of real-world broken objects and their complete counter- parts. In: CVPR. pp. 4681–4691 (2023).https://doi.org/10.1109/CVPR52729. 2023.00454

    2023

  8. [16]

    In: ICLR (2025)

    Lan, Y., Zhou, S., Lyu, Z., Hong, F., Yang, S., Dai, B., Pan, X., Loy, C.C.: Gaus- siananything: Interactive point cloud latent diffusion for 3d generation. In: ICLR (2025)

    2025

  9. [17]

    In: ICLR 2022 Workshop on Geometrical and Topological Representation Learning (GTRL) (2022),https: //openreview.net/forum?id=BY88eBbkpe5

    Leach, A., Schmon, S.M., Degiacomi, M.T., Willcocks, C.G.: Denoising diffusion probabilistic models on SO(3) for rotational alignment. In: ICLR 2022 Workshop on Geometrical and Topological Representation Learning (GTRL) (2022),https: //openreview.net/forum?id=BY88eBbkpe5

    2022

  10. [18]

    In: Proceedings of the 41st Inter- national Conference on Machine Learning

    Lee, N., Min, J., Lee, J., Kim, S., Lee, K., Park, J., Cho, M.: 3D geometric shape assembly via efficient point cloud matching. In: Proceedings of the 41st Inter- national Conference on Machine Learning. Proceedings of Machine Learning Re- search, vol. 235, pp. 26856–26873. PM...

    2024

  11. [19]

    Li, J., Cheng, C., Ma, J., Liu, G.: Geometric point attention transformer for 3d shape reassembly (2024), arXiv:2411.17788

    2024 arXiv

  12. [20]

    In: ICCV

    Li, S., Jiang, Z., Chen, G., Xu, C., Tan, S., Wang, X., Fang, I., Zyskowski, K., McPherron, S.P., Iovita, R., Feng, C., Zhang, J.: Garf: Learning generalizable 3d reassembly for real-world fractures. In: ICCV. pp. 5711–5721 (October 2025)

    2025

  13. [21]

    (2024).https://doi.org/10.52202/079017-3985

    Li, T., Katabi, D., He, K.: Return of unconditional generation: A self-supervised representationgenerationmethod.In:NeurIPS.vol.37,pp.125441–125468.Curran Associates, Inc. (2024).https://doi.org/10.52202/079017-3985

    2024 doi

  14. [22]

    In: NeurIPS (2025)

    Li, Z., Wang, Y., Zheng, H., Luo, Y., Wen, B.: Sparc3d: Sparse representation and construction for high-resolution 3d shapes modeling. In: NeurIPS (2025)

    2025

  15. [23]

    Liu, Q.: Rectified flow: A marginal preserving approach to optimal transport (2022), arXiv:2209.14577

    2022 arXiv

  16. [24]

    In: ICLR (2023),https://openreview.net/forum?id= XVjTT1nw5z

    Liu, X., Gong, C., Liu, Q.: Flow straight and fast: Learning to generate and transfer data with rectified flow. In: ICLR (2023),https://openreview.net/forum?id= XVjTT1nw5z

    2023

  17. [25]

    In: ICCV

    Lu, J., Hua, G., Huang, Q.: Jigsaw++: Imagining complete shape priors for object reassembly. In: ICCV. pp. 6704–6714 (October 2025)

    2025

  18. [26]

    Computer Graphics Forum44(2024)

    Lu, J., Liang, Y., Han, H., Hua, J., Jiang, J., Li, X., Huang, Q.X.: A survey on computational solutions for reconstructing complete objects by reassembling their fractured parts. Computer Graphics Forum44(2024)

    2024

  19. [27]

    In: NeurIPS

    Lu,J.,Sun,Y.,Huang,Q.:Jigsaw:Learningtoassemblemultiplefracturedobjects. In: NeurIPS. vol. 36, pp. 14969–14986. Curran Associates, Inc. (2023)

    2023

  20. [28]

    In: CVPR

    Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Re- paint: Inpainting using denoising diffusion probabilistic models. In: CVPR. pp. 11461–11471 (June 2022)

    2022

  21. [29]

    In: CVPR (2019)

    Mescheder, L., Oechsle, M., Niemeyer, M., Nowozin, S., Geiger, A.: Occupancy networks: Learning 3d reconstruction in function space. In: CVPR (2019)

    2019

  22. [30]

    Springer International Publishing (2020).https://doi.org/10.1007/978-3-030- 58452-8_24,https://doi.org/10.1007/978-3-030-58452-8_24

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.:Nerf:Representingscenesasneuralradiancefieldsforviewsynthesis.In:ECCV. Springer International Publishing (2020).https://doi.org/10.1007/978-3-030- 58452-8_24,https://doi.org/10.1007/978-3-030-5...

    2020 doi

  23. [31]

    arXiv preprint arXiv:2212.09748 (2022)

    Peebles, W., Xie, S.: Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748 (2022)

    2022 arXiv

  24. [32]

    In: International SARe: Structure-Aware Large-Scale 3D Fragment Reassembly 17 Conference on Learning Representations (ICLR) (2024),[https://openreview

    Perniás,P.,Rampas,D.,Richter,M.L.,Pal,C.J.,Aubreville,M.:Würstchen:Anef- ficient architecture for large-scale text-to-image diffusion models. In: International SARe: Structure-Aware Large-Scale 3D Fragment Reassembly 17 Conference on Learning Representations (ICLR) (2024),[htt...

    2024

  25. [33]

    In: CVPR

    Qin, Z., Yu, H., Wang, C., Guo, Y., Peng, Y., Xu, K.: Geometric transformer for fast and robust point cloud registration. In: CVPR. pp. 11133–11142 (2022). https://doi.org/10.1109/CVPR52688.2022.01086

    2022

  26. [34]

    In: CVPR

    Ren, X., Huang, J., Zeng, X., Museth, K., Fidler, S., Williams, F.: Xcube: Large- scale 3d generative modeling using sparse voxel hierarchies. In: CVPR. pp. 4209– 4219 (2024).https://doi.org/10.1109/CVPR52733.2024.00403

    2024

  27. [35]

    In: CVPR (June 2024)

    Scarpellini, G., Fiorini, S., Giuliari, F., Morerio, P., Del Bue, A.: Diffassemble: A unified graph-diffusion model for 2d and 3d reassembly. In: CVPR (June 2024)

    2024

  28. [36]

    In: NeurIPS

    Sellán, S., Chen, Y.C., Wu, Z., Garg, A., Jacobson, A.: Breaking bad: A dataset for geometric fracture and reassembly. In: NeurIPS. vol. 35, pp. 38885–38898 (2022)

    2022

  29. [37]

    ACM TOG42(1) (2023).https://doi.org/10.1145/3549540,https://doi.org/10.1145/3549540

    Sellán, S., Luong, J., Mattos Da Silva, L., Ramakrishnan, A., Yang, Y., Jacobson, A.: Breaking good: Fracture modes for realtime destruction. ACM TOG42(1) (2023).https://doi.org/10.1145/3549540,https://doi.org/10.1145/3549540

    2023 doi

  30. [38]

    In: ICML (2025)

    Shen, Y., Wu, R., Ke, Y., Song, X., Li, Z., Li, X., Fan, H., Lu, H., Dong, H.: Biassemble: Learning collaborative affordance for bimanual geometric assembly. In: ICML (2025)

    2025

  31. [39]

    In: NeurIPS (2025)

    Sun, T., Zhu, L., Huang, S., Song, S., Armeni, I.: Rectified point flow: Generic point cloud pose estimation. In: NeurIPS (2025)

    2025

  32. [40]

    In: NeurIPS

    Tancik, M., Srinivasan, P.P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Sing- hal, U., Ramamoorthi, R., Barron, J.T., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. In: NeurIPS. vol. 33 (2020)

    2020

  33. [41]

    In: Globerson, A., Mackey, L., Belgrave, D., Fan, A., Paquet, U., Tomczak, J., Zhang, C

    Tsesmelis, T., Palmieri, L., Khoroshiltseva, M., Islam, A., Elkin, G., Shahar, O.I., Scarpellini, G., Fiorini, S., Ohayon, Y., Alali, N., Aslan, S., Morerio, P., Vascon, S., Gravina, E., Napolitano, M.C., Scarpati, G., Zuchtriegel, G., Spühler, A., Fuchs, M.E., James, S., Ben-...

    2024

  34. [42]

    In: ICLR (2025)

    Wang,Z.,Chen,J.,Furukawa,Y.:Puzzlefusion++:Auto-agglomerative3dfracture assembly by denoise and verify. In: ICLR (2025)

    2025

  35. [43]

    In: ICCV

    Wu, R., Tie, C., Du, Y., Zhao, Y., Dong, H.: Leveraging se(3) equivariance for learning 3d geometric shape assembly. In: ICCV. pp. 14265–14274 (2023).https: //doi.org/10.1109/ICCV51070.2023.01316

    2023

  36. [44]

    In: CVPR

    Wu, T., Zhang, J., Fu, X., Wang, Y., Ren, J., Pan, L., Wu, W., Yang, L., Wang, J., Qian, C., Lin, D., Liu, Z.: Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In: CVPR. pp. 803–814 (2023). https://doi.org/10.1109/CVPR52...

    2023

  37. [45]

    In: CVPR

    Xiang, J., Lv, Z., Xu, S., Deng, Y., Wang, R., Zhang, B., Chen, D., Tong, X., Yang, J.: Structured 3d latents for scalable and versatile 3d generation. In: CVPR. pp. 21469–21480 (June 2025)

    2025

  38. [46]

    Journal of Com- putational Design and Engineering2(2), 79–87 (2015).https://doi.org/https: //doi.org/10.1016/j.jcde.2014.12.002

    Xu, J., Zhou, M., Wu, Z., Shui, W., Ali, S.: Robust surface segmentation and edge feature lines extraction from fractured fragments of relics. Journal of Com- putational Design and Engineering2(2), 79–87 (2015).https://doi.org/https: //doi.org/10.1016/j.jcde.2014.12.002

    2015 doi

  39. [47]

    In: SIGGRAPH Asia 2024 Confer- ence Papers

    Xu, Q.C., Chen, H.X., Hua, J., Zhan, X., Yang, Y.L., Mu, T.J.: Fragmentdiff: A diffusion model for fractured object assembly. In: SIGGRAPH Asia 2024 Confer- ence Papers. Association for Computing Machinery, New York, NY, USA (2024). https://doi.org/10.1145/3680528.3687673 18 H...

    2024

  40. [48]

    Xu, Z., Wang, J., Shi, Y., Gao, C., Zhou, G.: Topodiffuser: A diffusion-based multi- modal trajectory prediction model with topometric maps (2025), arXiv:2508.00303

    2025 arXiv

  41. [49]

    In: ICLR (2025)

    Yang, Y., Jiang, P.T., Hou, Q., Zhang, H., Chen, J., Li, B.: Multi-task dense predictions via unleashing the power of diffusion. In: ICLR (2025)

    2025

  42. [50]

    In: CVPR (2024)

    Ye, H., Xu, D.: Diffusionmtl: Learning multi-task denoising diffusion model from partially annotated data. In: CVPR (2024)

    2024

  43. [51]

    In: ICLR (2025),[https://openreview.net/forum?id=DJSZGGZYVi](https:// openreview.net/forum?id=DJSZGGZYVi)

    Yu, S., Kwak, S., Jang, H., Jeong, J., Huang, J., Shin, J., Xie, S.: Representation alignment for generation: Training diffusion transformers is easier than you think. In: ICLR (2025),[https://openreview.net/forum?id=DJSZGGZYVi](https:// openreview.net/forum?id=DJSZGGZYVi)

    2025

  44. [52]

    In: Koyejo, S., Mo- hamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A

    zeng, x., Vahdat, A., Williams, F., Gojcic, Z., Litany, O., Fidler, S., Kreis, K.: Lion: Latent point diffusion models for 3d shape generation. In: Koyejo, S., Mo- hamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) NeurIPS. vol. 35, pp. 10021–10039. Curran Associates...

    2022

  45. [53]

    ACM TOG42(4) (2023)

    Zhang, B., Tang, J., Nießner, M., Wonka, P.: 3dshape2vecset: A 3d shape represen- tation for neural fields and generative diffusion models. ACM TOG42(4) (2023). https://doi.org/10.1145/3592442

    2023 doi

  46. [54]

    In: ICLR (2025),https://openreview.net/forum?id= 72OSO38a2z

    Zhang, B., Wonka, P.: LaGeM: A large geometry model for 3d representation learning and diffusion. In: ICLR (2025),https://openreview.net/forum?id= 72OSO38a2z

    2025

  47. [55]

    In: Proceedings of the SIGGRAPH Asia 2025 Conference Papers

    Zhao, W., Cao, Y.P., Xu, J., Dong, Y., Shan, Y.: Assembler: Scalable 3d part as- sembly via anchor point diffusion. In: Proceedings of the SIGGRAPH Asia 2025 Conference Papers. SA Conference Papers ’25, Association for Computing Machin- ery (2025).https://doi.org/10.1145/37573...

    2025

  48. [56]

    In: CVPR

    Zhou, Y., Barnes, C., Lu, J., Yang, J., Li, H.: On the continuity of rotation representations in neural networks. In: CVPR. pp. 5738–5746 (2019).https: //doi.org/10.1109/CVPR.2019.00589

    2019

  49. [57]

    In: Digital Heritage

    Zuchtriegel, G., Brunetto, M.A., Gravina, E., Napolitano, M.C., Ricciardi, F.S., Zambrano, A., Khoroshiltseva, M., Palmieri, L., Pelillo, M., Vascon, S., Elkin, G., Shahar, O.I., Ohayon, Y., Alali, N., Ben-Shahar, O.: The repair project: Datasets for archaeological and restora...

    2025 doi

Pith tools

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