pith. sign in

arxiv: 2604.28134 · v2 · pith:Z76XCF5Ynew · submitted 2026-04-30 · 💻 cs.CV

MeshReGen: A Unified 3D Geometry Regeneration Framework

Geon Yeong Park , Roman Shapovalov , Rakesh Ranjan , Jong Chul Ye , Andrea Vedaldi , Thu Nguyen-Phuoc This is my paper

Pith reviewed 2026-05-21 00:44 UTC · model grok-4.3

classification 💻 cs.CV
keywords meshregentasksallowsfine-grainedgeometryimagesinitialregeneration
0
0 comments X

The pith

MeshReGen regenerates 3D objects from initial shapes and images using a VecSet conditioning mechanism.

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

MeshReGen is presented as a 3D regenerator that takes an initial 3D shape as input rather than generating objects in a one-shot manner from text or images alone. This allows it to support tasks such as 3D enhancement, reconstruction from images, and editing by updating the input geometry with fine details. The approach learns a regeneration prior through self-supervised pretext tasks and augmentations on standard 3D datasets without needing extra labels. A sympathetic reader would care because this shifts 3D generation toward more controllable and consistent outputs across multiple applications.

Core claim

MeshReGen conditions a 3D regenerator on an initial 3D shape and employs a new VecSet-based conditioning mechanism to update or improve the input geometry with consistent fine-grained details. It learns a widely applicable regeneration prior from off-the-shelf 3D datasets via self-supervised pretext tasks and augmentations without additional annotations, and achieves state-of-the-art performance in controllable 3D generation across several tasks including enhancement, reconstruction, and editing.

What carries the argument

VecSet conditioning mechanism that enables the regenerator to refine the input 3D shape with consistent details.

Load-bearing premise

That a conceptually simple conditioning on an initial 3D shape together with VecSet will support many tasks and allow learning a general regeneration prior from self-supervised pretext tasks on off-the-shelf datasets.

What would settle it

Observing that the output meshes fail to maintain geometric consistency with the input shape or do not improve fine-grained quality on standard evaluation benchmarks would falsify the effectiveness of the conditioning approach.

Figures

Figures reproduced from arXiv: 2604.28134 by Andrea Vedaldi, Geon Yeong Park, Jong Chul Ye, Rakesh Ranjan, Roman Shapovalov, Thu Nguyen-Phuoc.

Figure 1
Figure 1. Figure 1: 3D-ReGen is a diffusion-based 3D regeneration framework that reconstructs complete 3D shapes from coarse geometry using 2D image cues. The regeneration prior is learned through self-supervised pretext tasks and augmentations, without task￾specific architectures or extra annotations. Users can input a coarse mesh (3D Enhance￾ment), incomplete point cloud (3D from Point Cloud), or masked mesh (3D Editing) to… view at source ↗
Figure 2
Figure 2. Figure 2: Overview. 3D-ReGen takes both 2D image and initial 3D geometry as in￾put, enabling explicit control over global geometry (e.g., pose, coarse shape) while improving fine-grained details. The 3D condition is encoded as VecSet latents (z L ) that compactly represent global geometry. After summing with positional embeddings, these conditionings and random latents are diffused by a DiT into enhanced latents, th… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative results of object-level 3D scene refinement. (Top) A single-scene image is converted into a coarse monolithic 3D scene by Sparc3D [31], where details of each object are degraded due to the limited capacity of the latent code. To mitigate this degradation, we segment the scene with AutoPartGen [5], refine individual objects with 3D-ReGen (Bottom), and update each object to produce a high-quality… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison with 3D generation and enhancement baselines. Coarse 3D conditions are obtained as parts of the original scene [32], and 2D image conditions are automatically regenerated given initial renders of each coarse object, as detailed in Sec. 4.1. Using these multi-modal conditions, 3D-ReGen regenerates the fine-grained details of 3D shapes, outperforming the recent enhancement baseline [11… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparisons for faithful Image-to-3D generation on the GSO dataset [12]. 3D-ReGen preserves the geometry of the initial VGGT point cloud [58] while producing clean surfaces. Baselines that rely solely on image conditioning tend to be less faithful and smooth out fine details. For VGGT and 3D-ReGen, the number of input views is shown on the left; for the baselines, it is shown in parentheses. us… view at source ↗
Figure 6
Figure 6. Figure 6: Reconstruction results of VGGT+3D-ReGen from one and two images. Note the improvement in the details on the side that is occluded in the first view. The trend continues with more views, as shown by metrics in Tab. 2. the different tasks one wishes to solve (e.g., enhancement, reconstruction, edit￾ing). 3D training data is already scarce, and procuring paired data for each task is much harder still. Hence, … view at source ↗
Figure 7
Figure 7. Figure 7: 3D object editing examples. Original shapes and edited images (shown in inset) are from [26]. The resulting shape maintains both the orientation and the level of detail of the input shape. The 3D edit M masking the edit region is also visualized. GSO dataset [12]. For perceptual evaluation, we compute PSNR, SSIM [60], and LPIPS [84] between the rendered and ground-truth normal maps by rescaling unit normal… view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparisons of object editing with [26]. Red boxes highlight either artifacts present in the editing results or differences between the edited shape and the original shape. The 3D edit M masking the edit region is also visualized. Input Views 4 views 3 views TRELLIS TRELLIS + VGGT Ours (a) Reconstruction (b) Enhancement Cond. Image Coarse Input + TRELLIS + Ours view at source ↗
Figure 9
Figure 9. Figure 9: TRELLIS results on regeneration tasks. (a) Reconstruction from sparse views. We treat VGGT predictions as sparse structures in TRELLIS’s first stage (TRELLIS + VGGT). (b) Enhancement from degraded geometry. TRELLIS struggles to correct artifacts. Both TRELLIS and ours use exactly the same input shape or VGGT predic￾tions view at source ↗
Figure 10
Figure 10. Figure 10 view at source ↗
Figure 11
Figure 11. Figure 11: Data construction pipeline for 3D shape enhancement. We synthesize com￾positional scenes by placing x H on a multi-object grid (3 × 3 shown). The scene is encoded into VecSet tokens z0, noised at t ∗ = 350 (under a variance-preserving (VP) schedule [16, 51]), and denoised to t = 0 using a pre-trained 3D generator and the render of the grid as image conditioning. Degraded shapes can then be extracted as cl… view at source ↗
Figure 12
Figure 12. Figure 12: Paired examples of ground-truth meshes x H and their degraded counterparts. Each x H is randomly placed in a predefined 3 × 3 grid within the unit cube [−1, 1]3 . into VecSet tokens. Following the sampling process of compositional 3D gener￾ators, we perturb these tokens at a moderate noise level t ∗ = 350 < T inspired by SDEdit [41]. The noisy scene tokens are then denoised using a corresponding view at source ↗
Figure 13
Figure 13. Figure 13: Histogram of F-score@1% over ground-truth and degraded meshes. Simple manual degradations, such as low-pass fil￾tering of the SDF or Taubin smoothing [54], do not allow controlling the degradation consistently across objects with different geometric complexity, as similarly reported in [10]. Analysis of degraded datasets. Using F-score@1%, view at source ↗
Figure 14
Figure 14. Figure 14: The reference scene image and the top-down render of the generated recon￾struction, with the target object highlighted in red, are provided to the VLM to leverage scene context. The VLM then refines the coarse object rendering to generate the final complete, detailed, and enhanced image of the object view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative results of 3D shape enhancement on the held-out validation set. 3D-ReGen demonstrates its robustness, enhancing the shape even when both image and shape conditions are degraded. Refining lower-quality 3D assets We present preliminary results showing that 3D-ReGen can enhance low-quality 3D assets arising from noisy or low-resolution scans, compression artifacts, legacy content, degraded scene … view at source ↗
Figure 16
Figure 16. Figure 16: Qualitative results for 3D-ReGen on 3D asset enhancement. 3D-ReGen shows promising improvements in recovering fine details and refining artifacts in the original generated shapes, resulting in outputs that better match the input image. Cond. Image Input block-out + CLAY v1.5 + CLAY v2.0 + Ours view at source ↗
Figure 17
Figure 17. Figure 17: Qualitative results of 3D-ReGen on 3D shape regeneration from block-outs. Compared to CLAY, 3D-ReGen is capable of adding fine details while preserving the underlying shapes and proportions of the block-outs. train for 25K steps on 3 × 3 degraded shapes, while those in Tab. 4 use C = 512 and train for 25K steps on 2 × 2 degraded shapes view at source ↗
Figure 18
Figure 18. Figure 18: More qualitative examples from the ablation study. (a) Samples from the ablation study of different diffusion parameterizations in Tab. 4. (b) Samples from the ablation study of the conditioning mechanism in Tab. 3. The proposed method consistently outperforms other design choices with fine-grained details view at source ↗
Figure 19
Figure 19. Figure 19: Additional qualitative comparisons against 3D-enhancement baselines. We obtain coarse 3D conditions by extracting parts from the original scene [32]. Leverag￾ing coarse 3D shapes and guide images, 3D-ReGen regenerates fine-grained 3D shape details and outperforms the state-of-the-art point-cloud-based method for conditional 3D generation, CLAY [81]. (c) Architecture. Hunyuan3D-Omni converts all conditioni… view at source ↗
Figure 20
Figure 20. Figure 20: Qualitative comparison of two-view object reconstruction with Hunyuan3D￾Omni [19] and CLAY [81]. 3D-ReGen produces more complete, faithful and higher￾fidelity shapes. erative model that directly produces SDF representations. Leveraging a versatile control space and self-supervision with base 3D datasets, our approach enables a controllable 3D generation pipeline supporting diverse downstream tasks, in￾clu… view at source ↗
read the original abstract

We consider the problem of regenerating 3D objects from 2D images and initial 3D shapes. Most 3D generators operate in a one-shot fashion, converting text or images to a 3D object with limited controllability. We introduce instead MeshReGen, a 3D regenerator that is conditioned on an initial 3D shape. This conceptually simple formulation allows us to support numerous useful tasks, including 3D enhancement, reconstruction, and editing. MeshReGen uses a new conditioning mechanism based on VecSet, which allows the regenerator to update or improve the input geometry with consistent fine-grained details. MeshReGen learns a widely applicable regeneration prior from off-the-shelf 3D datasets via self-supervised pretext tasks and augmentations, without additional annotations. We evaluate both the geometric consistency and fine-grained quality of MeshReGen, achieving state-of-the-art performance in controllable 3D generation across several tasks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces MeshReGen, a unified 3D geometry regeneration framework conditioned on an initial 3D shape. It proposes a VecSet-based conditioning mechanism to enable updating the input geometry while adding consistent fine-grained details. The model learns a regeneration prior via self-supervised pretext tasks and augmentations on off-the-shelf 3D datasets without extra annotations. This formulation supports multiple tasks including 3D enhancement, reconstruction, and editing. The authors evaluate geometric consistency and fine-grained quality, claiming state-of-the-art performance in controllable 3D generation across several tasks.

Significance. If the results hold, the work offers a conceptually simple unified approach to controllable 3D tasks that could reduce reliance on separate models for enhancement, reconstruction, and editing. The self-supervised training on existing datasets without annotations is a clear strength, as is the focus on geometric consistency through conditioning. This could advance practical 3D generation pipelines in computer vision and graphics.

major comments (2)
  1. [§3.3] §3.3: The self-supervised loss relies on reconstruction from pretext tasks and augmentations. Without an explicit term penalizing deviation from the input geometry on unchanged regions (e.g., a masked consistency loss between input and regenerated mesh), the model could alter input structure arbitrarily, undermining the central claim of preserving structure while adding consistent fine-grained details via VecSet conditioning.
  2. [§4.1, Table 2] §4.1, Table 2: The SOTA claims for controllable generation rest on quantitative comparisons, but the absence of error bars, statistical tests, or detailed ablation on the VecSet component versus standard conditioning makes it difficult to confirm that the improvements are robust and attributable to the proposed mechanism rather than dataset or training specifics.
minor comments (2)
  1. [§3] The VecSet conditioning is described at a high level in the method; adding a precise equation or pseudocode for the conditioning operation early in §3 would improve verifiability.
  2. [Figures] Figure captions for qualitative results could more explicitly label input mesh, regenerated output, and reference for each task to aid reader interpretation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback on our manuscript. We have addressed each of the major comments in detail below and made corresponding revisions to the paper to enhance its clarity and the robustness of the presented results.

read point-by-point responses

  1. Referee: [§3.3] §3.3: The self-supervised loss relies on reconstruction from pretext tasks and augmentations. Without an explicit term penalizing deviation from the input geometry on unchanged regions (e.g., a masked consistency loss between input and regenerated mesh), the model could alter input structure arbitrarily, undermining the central claim of preserving structure while adding consistent fine-grained details via VecSet conditioning.

    Authors: We appreciate the referee's observation regarding the loss formulation. The VecSet conditioning is specifically engineered to enable localized updates by modeling geometric differences through its set-based attention, and the self-supervised pretext tasks (including reconstruction from augmented inputs) are designed to encourage fidelity to the original structure. Nevertheless, we acknowledge that an explicit term could provide additional safeguards. In the revised manuscript we have added a masked consistency loss that penalizes deviations exclusively on regions outside the augmentation masks. This term is integrated into the overall objective in Section 3.3, and we report its effect in a new ablation in the supplementary material. The core self-supervised framework remains unchanged. revision: yes

  2. Referee: [§4.1, Table 2] §4.1, Table 2: The SOTA claims for controllable generation rest on quantitative comparisons, but the absence of error bars, statistical tests, or detailed ablation on the VecSet component versus standard conditioning makes it difficult to confirm that the improvements are robust and attributable to the proposed mechanism rather than dataset or training specifics.

    Authors: We agree that error bars, statistical significance testing, and a targeted ablation would strengthen the quantitative claims. In the revised manuscript we have rerun the main experiments across five random seeds and added standard deviation error bars to Table 2. We have also inserted a new ablation subsection in §4.1 that directly compares VecSet conditioning against a standard concatenation baseline under identical training settings, confirming that the observed gains are attributable to the VecSet mechanism. Finally, we include paired t-test p-values to establish statistical significance of the improvements over prior methods. revision: yes

Circularity Check

0 steps flagged

No circularity: regeneration prior learned from external off-the-shelf data via self-supervision

full rationale

The abstract and provided text present MeshReGen as learning a regeneration prior directly from external 3D datasets using self-supervised pretext tasks and augmentations, without additional annotations. The VecSet conditioning is introduced as a new mechanism to support tasks like enhancement and editing, but no equations, fitted parameters, or self-citations are shown that reduce the central claims (consistent fine-grained details, unified tasks) to inputs defined inside the work. The derivation remains self-contained against external benchmarks, with no load-bearing steps that equate predictions to constructions or prior self-work.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The claim rests on the effectiveness of VecSet as a conditioning mechanism and on the domain assumption that self-supervised pretext tasks suffice to learn a regeneration prior usable across multiple downstream tasks. VecSet is introduced as a new entity without independent evidence outside the paper.

axioms (1)
  • domain assumption Self-supervised pretext tasks and augmentations on off-the-shelf 3D datasets suffice to learn a widely applicable regeneration prior without additional annotations.
    Invoked in the abstract to justify training the regenerator for enhancement, reconstruction, and editing.
invented entities (1)
  • VecSet no independent evidence
    purpose: New conditioning mechanism that updates or improves input geometry while preserving consistent fine-grained details.
    Introduced in the abstract as the core technical novelty; no external falsifiable evidence is referenced.

pith-pipeline@v0.9.0 · 5710 in / 1402 out tokens · 42644 ms · 2026-05-21T00:44:33.263670+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

80 extracted references · 80 canonical work pages · 14 internal anchors

  1. [1]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Barda, A., Gadelha, M., Kim, V.G., Aigerman, N., Bermano, A.H., Groueix, T.: Instant3dit: Multiview inpainting for fast editing of 3d objects. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 16273–16282 (2025) 13, 11

    work page 2025

  2. [2]

    arXiv preprint arXiv:2403.12032 (2024) 13, 11

    Chen, H., Shi, R., Liu, Y., Shen, B., Gu, J., Wetzstein, G., Su, H., Guibas, L.: Generic 3d diffusion adapter using controlled multi-view editing. arXiv preprint arXiv:2403.12032 (2024) 13, 11

    work page arXiv 2024

  3. [3]

    In: Proceedings of the European Conference on Computer Vi- sion (ECCV) (2024) 11

    Chen, M., Laina, I., Vedaldi, A.: DGE: Direct gaussian 3D editing by consistent multi-view editing. In: Proceedings of the European Conference on Computer Vi- sion (ECCV) (2024) 11

    work page 2024

  4. [4]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2025) 8

    Chen, M., Shapovalov, R., Laina, I., Monnier, T., Wang, J., Novotny, D., Vedaldi, A.: PartGen: Part-level 3D generation and reconstruction with multi-view diffusion models. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2025) 8

    work page 2025

  5. [5]

    Autopartgen: Autogres- sive 3d part generation and discovery.arXiv preprint arXiv:2507.13346, 2025

    Chen, M., Wang, J., Shapovalov, R., Monnier, T., Jung, H., Wang, D., Ranjan, R., Laina, I., Vedaldi, A.: Autopartgen: Autogressive 3d part generation and discovery. arXiv preprint arXiv:2507.13346 (2025) 5, 8, 10

    work page arXiv 2025

  6. [6]

    In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2025) 7

    Chen, M., Wang, J., Shapovalov, R., Monnier, T., Jung, H., Wang, D., Ranjan, R., Laina, I., Vedaldi, A.: AutoPartGen: Autogressive 3D part generation and discovery. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2025) 7

    work page 2025

  7. [7]

    Dora: Sampling and benchmarking for 3d shape variational auto-encoders

    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. arXiv2412.17808(2024) 4

    work page arXiv 2024

  8. [8]

    Objaverse-XL: A Universe of 10M+ 3D Objects

    Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., VanderBilt, E., Kembhavi, A., Vondrick, C., Gkioxari, G., Ehsani, K., Schmidt, L., Farhadi, A.: Objaverse-XL: A universe of 10M+ 3D objects. CoRRabs/2307.05663(2023) 4

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  9. [9]

    In: Proc

    Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objaverse: A universe of annotated 3D objects. In: Proc. CVPR (2023) 4, 8

    work page 2023

  10. [11]

    arXiv2411.16820(2024) 3, 4, 6, 13, 11

    Deng, K., Guo, Y., Sun, J., Zou, Z., Li, Y., Cai, X., Cao, Y., Liu, Y., Liang, D.: DetailGen3D: generative 3D geometry enhancement via data-dependent flow. arXiv2411.16820(2024) 3, 4, 6, 13, 11

    work page arXiv 2024

  11. [12]

    In: Proc

    Downs, L., Francis, A., Koenig, N., Kinman, B., Hickman, R., Reymann, K., McHugh, T.B., Vanhoucke, V.: Google Scanned Objects: A high-quality dataset of 3D scanned household items. In: Proc. ICRA (2022) 7, 11, 12, 8, 10

    work page 2022

  12. [13]

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

    Gao, W., Wang, D., Fan, Y., Bozic, A., Stuyck, T., Li, Z., Dong, Z., Ranjan, R., Sarafianos, N.: 3D mesh editing using masked LRMs. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 7154–7165 (2025) 9

    work page 2025

  13. [14]

    The Llama 3 Herd of Models

    Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Let- man, A., Mathur, A., Schelten, A., Vaughan, A., et al.: The llama 3 herd of models. arXiv preprint arXiv:2407.21783 (2024) 11

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  14. [15]

    Journal of the Engineering Mechanics Division102(5), 749–756 (1976) 9, 3

    Herrmann, L.R.: Laplacian-isoparametric grid generation scheme. Journal of the Engineering Mechanics Division102(5), 749–756 (1976) 9, 3

    work page 1976

  15. [16]

    Advances in Neural Information Processing Systems33, 6840–6851 (2020) 2

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems33, 6840–6851 (2020) 2

    work page 2020

  16. [17]

    In: Proc

    Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: LRM: Large reconstruction model for single image to 3D. In: Proc. ICLR (2024) 11

    work page 2024

  17. [18]

    Hunyuan3D 2.1: From Images to High-Fidelity 3D Assets with Production-Ready PBR Material

    Hunyuan3D, T., Yang, S., Yang, M., Feng, Y., Huang, X., Zhang, S., He, Z., Luo, D., Liu, H., Zhao, Y., Lin, Q., Lai, Z., Yang, X., Shi, H., Zhao, Z., Zhang, B., Yan, H., Wang, L., Liu, S., Zhang, J., Chen, M., Dong, L., Jia, Y., Cai, Y., Yu, J., Tang, Y., Guo, D., Yu, J., Zhang, H., Ye, Z., He, P., Wu, R., Wei, S., Zhang, C., Tan, Y., Sun, Y., Niu, L., Hu...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  18. [19]

    arXiv preprint arXiv:2509.21245 (2025) 3, 4, 12, 14, 8, 11

    Hunyuan3D, T., Zhang, B., Guo, C., Liu, H., Yan, H., Shi, H., Huang, J., Yu, J., Li, K., Wang, P., et al.: Hunyuan3d-omni: A unified framework for controllable generation of 3d assets. arXiv preprint arXiv:2509.21245 (2025) 3, 4, 12, 14, 8, 11

    work page arXiv 2025

  19. [20]

    In: Proc

    Jin, H., Jiang, H., Tan, H., Zhang, K., Bi, S., Zhang, T., Luan, F., Snavely, N., Xu, Z.: LVSM: a large view synthesis model with minimal 3D inductive bias. In: Proc. ICLR (2025) 11

    work page 2025

  20. [21]

    In: Proceed- ings of the fourth Eurographics symposium on Geometry processing (2006) 11, 12

    Kazhdan, M., Bolitho, M., Hoppe, H.: Poisson surface reconstruction. In: Proceed- ings of the fourth Eurographics symposium on Geometry processing (2006) 11, 12

    work page 2006

  21. [22]

    MapAnything: Universal Feed-Forward Metric 3D Reconstruction

    Keetha, N., Müller, N., Schönberger, J., Porzi, L., Zhang, Y., Fischer, T., Knapitsch, A., Zauss, D., Weber, E., Antunes, N., Luiten, J., Lopez-Antequera, M., Bulò, S.R., Richardt, C., Ramanan, D., Scherer, S., Kontschieder, P.: Ma- pAnything: universal feed-forward metric 3D reconstruction. arXiv2509.13414 (2025) 9

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  22. [23]

    FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space

    Labs,B.F.,Batifol,S.,Blattmann,A.,Boesel,F.,Consul,S.,Diagne,C.,Dockhorn, T., English, J., English, Z., Esser, P., Kulal, S., Lacey, K., Levi, Y., Li, C., Lorenz, D., Müller, J., Podell, D., Rombach, R., Saini, H., Sauer, A., Smith, L.: FLUX.1 kontext: Flow matching for in-context image generation and editing in latent space. arXiv2506.15742(2025) 4, 9, 13

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  23. [24]

    arXiv2509.12815(2025) 4

    Lei, B., Li, Y., Liu, X., Yang, S., Xu, L., Huang, J., Tang, R., Weng, H., Liu, J., Xu, J., Zhou, Z., Zhu, Y., Xing, J., Xu, J., Ma, C., Yan, X., Yang, Y., Wang, C., Xu, D., Ma, X., Chen, Y., Li, J., Yang, M., Zhang, S., Feng, Y., Huang, X., Luo, 3D-ReGen: A Unified 3D Geometry Regeneration Framework 17 D., He, Z., Jiang, P., Hu, C., Qin, Z., Miao, S., Li...

    work page arXiv 2025

  24. [26]

    arXiv preprint arXiv:2508.19247 , year=

    Li, L., Huang, Z., Feng, H., Zhuang, G., Chen, R., Guo, C., Sheng, L.: Voxhammer: Training-free precise and coherent 3d editing in native 3d space. arXiv preprint arXiv:2508.19247 (2025) 9, 12, 13, 11

    work page arXiv 2025

  25. [28]

    Li, W., Liu, J., Yan, H., Chen, R., Liang, Y., Chen, X., Tan, P., Long, X.: Crafts- man3d: High-fidelity mesh generation with 3d native generation and interactive geometry refiner (2024) 4

    work page 2024

  26. [29]

    2025.doi:10.48550/arXiv.2405.14979

    Li, W., Liu, J., Yan, H., Chen, R., Liang, Y., Chen, X., Tan, P., Long, X.: Crafts- man3d: High-fidelity mesh generation with 3d native generation and interactive geometry refiner. arXiv preprint arXiv:2405.14979 (2024) 11

    work page arXiv 2024

  27. [30]

    TripoSG: High-Fidelity 3D Shape Synthesis using Large-Scale Rectified Flow Models

    Li, Y., Zou, Z.X., Liu, Z., Wang, D., Liang, Y., Yu, Z., Liu, X., Guo, Y.C., Liang, D., Ouyang, W., Cao, Y.P.: TripoSG: high-fidelity 3D shape synthesis using large- scale rectified flow models. arXiv2502.06608(2025) 5, 6, 10

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  28. [32]

    Sparc: Sparse representation and construc- tion for high-resolution 3d shapes modeling.arXiv preprint arXiv:2505.14521, 2025

    Li, Z., Wang, Y., Zheng, H., Luo, Y., Wen, B.: Sparc3D: Sparse representation and construction for high-resolution 3d shapes modeling. arXiv2505.14521(2025) 6, 8, 10, 11

    work page arXiv 2025

  29. [33]

    Depth Anything 3: Recovering the Visual Space from Any Views

    Lin, H., Chen, S., Liew, J., Chen, D.Y., Li, Z., Shi, G., Feng, J., Kang, B.: Depth anything 3: Recovering the visual space from any views. arXiv2511.10647(2025) 9

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  30. [34]

    Flow matching for generative modeling.arXiv preprint arXiv:2305.08891, 2023

    Lin, S., Liu, B., Li, J., Yang, X.: Common diffusion noise schedules and sample steps are flawed. arXiv.csabs/2305.08891(2023) 1

    work page arXiv 2023

  31. [35]

    In: The Eleventh International Conference on Learning Representations (2023),https://openreview.net/forum?id=PqvMRDCJT9t14

    Lipman, Y., Chen, R.T.Q., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. In: The Eleventh International Conference on Learning Representations (2023),https://openreview.net/forum?id=PqvMRDCJT9t14

    work page 2023

  32. [36]

    In: Proc

    Liu, A., Lin, C., Liu, Y., Long, X., Dou, Z., Guo, H.X., Luo, P., Wang, W.: Part123: Part-aware 3D reconstruction from a single-view image. In: Proc. SIGGRAPH (2024) 8

    work page 2024

  33. [37]

    In: European Conference on Computer Vision

    Liu, F., Wang, H., Chen, W., Sun, H., Duan, Y.: Make-your-3d: Fast and consis- tent subject-driven 3d content generation. In: European Conference on Computer Vision. pp. 389–406. Springer (2024) 13, 11

    work page 2024

  34. [38]

    In: The Eleventh International Conference on Learning Representations (2023),https://openreview.net/forum?id=XVjTT1nw5z14 18 G

    Liu, X., Gong, C., qiang liu: Flow straight and fast: Learning to generate and trans- fer data with rectified flow. In: The Eleventh International Conference on Learning Representations (2023),https://openreview.net/forum?id=XVjTT1nw5z14 18 G. Y. Park et al

    work page 2023

  35. [39]

    ACM Computer Graphocs21(24) (1987) 4

    Lorensen, W., Cline, H.: Marching cubes: A high resolution 3D surface construction algorithm. ACM Computer Graphocs21(24) (1987) 4

    work page 1987

  36. [40]

    In: Proc

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: Proc. ICLR (2019) 1

    work page 2019

  37. [41]

    In: Proc

    Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: SDEdit: guided image synthesis and editing with stochastic differential equations. In: Proc. ICLR (2022) 8, 2

    work page 2022

  38. [42]

    In: Proceedings of the nineteenth annual symposium on Computational geometry

    Mitra, N.J., Nguyen, A.: Estimating surface normals in noisy point cloud data. In: Proceedings of the nineteenth annual symposium on Computational geometry. pp. 322–328 (2003) 9

    work page 2003

  39. [43]

    DINOv2: Learning Robust Visual Features without Supervision

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023) 6

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  40. [44]

    In: Proc

    Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proc. ICCV (2023) 3, 6

    work page 2023

  41. [45]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10684–10695 (2022) 1

    work page 2022

  42. [46]

    In: Proc

    Salimans, T., Ho, J.: Progressive distillation for fast sampling of diffusion models. In: Proc. ICLR (2022) 10

    work page 2022

  43. [47]

    In: Proceedings of the 19th annual conference on Computer graphics and interactive techniques

    Schroeder, W.J., Zarge, J.A., Lorensen, W.E.: Decimation of triangle meshes. In: Proceedings of the 19th annual conference on Computer graphics and interactive techniques. pp. 65–70 (1992) 9, 3

    work page 1992

  44. [48]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Sella, E., Fiebelman, G., Hedman, P., Averbuch-Elor, H.: Vox-e: Text-guided voxel editing of 3d objects. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 430–440 (2023) 13, 11

    work page 2023

  45. [49]

    In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2024) 4, 11, 12

    Siddiqui, Y., Kokkinos, F., Monnier, T., Kariya, M., Kleiman, Y., Garreau, E., Gafni, O., Neverova, N., Vedaldi, A., Shapovalov, R., Novotny, D.: Meta 3D As- set Gen: Text-to-mesh generation with high-quality geometry, texture, and PBR materials. In: Proceedings of Advances in Neural Information Processing Systems (NeurIPS) (2024) 4, 11, 12

    work page 2024

  46. [50]

    Advances in Neural Information Processing Systems37, 9532–9564 (2024) 11

    Siddiqui, Y., Monnier, T., Kokkinos, F., Kariya, M., Kleiman, Y., Garreau, E., Gafni, O., Neverova, N., Vedaldi, A., Shapovalov, R., et al.: Meta 3d assetgen: Text-to-mesh generation with high-quality geometry, texture, and pbr materials. Advances in Neural Information Processing Systems37, 9532–9564 (2024) 11

    work page 2024

  47. [51]

    Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score- basedgenerativemodelingthroughstochasticdifferentialequations.In:Proc.ICLR (2021) 2

    work page 2021

  48. [52]

    In: Proceedings of the 2004 Eurographics/ACM SIGGRAPH symposium on Geometry processing

    Sorkine, O., Cohen-Or, D., Lipman, Y., Alexa, M., Rössl, C., Seidel, H.P.: Lapla- cian surface editing. In: Proceedings of the 2004 Eurographics/ACM SIGGRAPH symposium on Geometry processing. pp. 175–184 (2004) 9, 3

    work page 2004

  49. [53]

    Lgm: Large multi-view gaussian model for high-resolution 3d content creation.arXiv preprint arXiv:2402.05054, 2024

    Tang, J., Chen, Z., Chen, X., Wang, T., Zeng, G., Liu, Z.: LGM: Large multi-view Gaussian model for high-resolution 3D content creation. arXiv2402.05054(2024) 11

    work page arXiv 2024

  50. [54]

    In: Proceedings of IEEE international conference on computer vision

    Taubin, G.: Curve and surface smoothing without shrinkage. In: Proceedings of IEEE international conference on computer vision. pp. 852–857. IEEE (1995) 3

    work page 1995

  51. [55]

    Gemini: A Family of Highly Capable Multimodal Models

    Team, G., Anil, R., Borgeaud, S., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., Millican, K., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023) 4 3D-ReGen: A Unified 3D Geometry Regeneration Framework 19

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  52. [56]

    TripoAI: Tripo3D text-to-3D (2024),https://www.tripo3d.ai2, 4, 10

    work page 2024

  53. [57]

    In: Proc

    Wang, D., Jung, H., Monnier, T., Sohn, K., Zou, C., Xiang, X., Yeh, Y.Y., Liu, D., Huang, Z., Nguyen-Phuoc, T., Fan, Y., Oprea, S., Wang, Z., Shapovalov, R., Sarafianos, N., Groueix, T., Toisoul, A., Dhar, P., Chu, X., Chen, M., Park, G.Y., Ranjan, R., Vedaldi, A.: WorldGen: From text to traversable and interactive 3D worlds. In: Proc. CVPR (2026) 8

    work page 2026

  54. [58]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2025) 7, 9

    Wang, J., Chen, M., Karaev, N., Vedaldi, A., Rupprecht, C., Novotny, D.: VGGT: Visual geometry grounded transformer. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2025) 7, 9

    work page 2025

  55. [59]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Wang, J., Chen, M., Karaev, N., Vedaldi, A., Rupprecht, C., Novotny, D.: Vggt: Visual geometry grounded transformer. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 5294–5306 (2025) 11

    work page 2025

  56. [60]

    In: Proc

    Wang, J., Thiesson, B., Xu, Y., Cohen, M.F.: Image and video segmentation by anisotropic kernel mean shift. In: Proc. ECCV (2004) 12

    work page 2004

  57. [61]

    $\pi^3$: Permutation-Equivariant Visual Geometry Learning

    Wang, Y., Zhou, J., Zhu, H., Chang, W., Zhou, Y., Li, Z., Chen, J., Pang, J., Shen, C., He, T.:π 3: Permutation-equivariant visual geometry learning. arXiv 2507.13347(2025) 9

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  58. [63]

    Meshlrm: Large reconstruction model for high- quality mesh

    Wei, X., Zhang, K., Bi, S., Tan, H., Luan, F., Deschaintre, V., Sunkavalli, K., Su, H., Xu, Z.: Meshlrm: Large reconstruction model for high-quality meshes. arXiv preprint arXiv:2404.12385 (2024) 11, 12

    work page arXiv 2024

  59. [64]

    Qwen-Image Technical Report

    Wu, C., Li, J., Zhou, J., Lin, J., Gao, K., Yan, K., Yin, S.m., Bai, S., Xu, X., Chen, Y., et al.: Qwen-image technical report. arXiv preprint arXiv:2508.02324 (2025) 4

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  60. [66]

    Structured 3D Latents for Scalable and Versatile 3D Generation

    Xiang, J., Lv, Z., Xu, S., Deng, Y., Wang, R., Zhang, B., Chen, D., Tong, X., Yang, J.: Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506 (2024) 4

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  61. [67]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    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: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 21469–21480 (2025) 11, 12, 15

    work page 2025

  62. [69]

    InstantMesh: Efficient 3D Mesh Generation from a Single Image with Sparse-view Large Reconstruction Models

    Xu, J., Cheng, W., Gao, Y., Wang, X., Gao, S., Shan, Y.: Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191 (2024) 11, 12

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  63. [70]

    Advances in Neural Information Processing Systems36, 15903–15935 (2023) 11

    Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagere- ward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems36, 15903–15935 (2023) 11

    work page 2023

  64. [71]

    Advances in Neural Information Processing Systems36(2024) 11

    Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagere- ward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems36(2024) 11

    work page 2024

  65. [72]

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

    Xue, L., Gao, M., Xing, C., Martín-Martín, R., Wu, J., Xiong, C., Xu, R., Niebles, J.C., Savarese, S.: Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1179–1189 (2023) 11 20 G. Y. Park et al

    work page 2023

  66. [73]

    arXiv preprint arXiv:2506.21076 (2025) 4, 8

    Yan, H., Luo, K., Li, W., Liang, Y., Li, S., Huang, J., Guo, C., Tan, P.: Posemaster: Generating 3d characters in arbitrary poses from a single image. arXiv preprint arXiv:2506.21076 (2025) 4, 8

    work page arXiv 2025

  67. [74]

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

    Yang, Y., Guo, Y.C., Huang, Y., Zou, Z.X., Yu, Z., Li, Y., Cao, Y.P., Liu, X.: HoloPart: generative 3d part amodal segmentation. arXiv2504.07943(2025) 8

    work page arXiv 2025

  68. [75]

    ACM Trans- actions on Graphics (TOG)43(6), 1–18 (2024) 11

    Ye, C., Qiu, L., Gu, X., Zuo, Q., Wu, Y., Dong, Z., Bo, L., Xiu, Y., Han, X.: Sta- blenormal: Reducing diffusion variance for stable and sharp normal. ACM Trans- actions on Graphics (TOG)43(6), 1–18 (2024) 11

    work page 2024

  69. [76]

    arXiv preprint (2025) 15

    Yenphraphai, J., Mirzaei, A., Chen, J., Zou, J., Tulyakov, S., Yeh, R.A., Wonka, P., Wang, C.: Shapegen4d: Towards high quality 4d shape generation from videos. arXiv preprint (2025) 15

    work page 2025

  70. [77]

    Qwen-image-layered: Towards inherent editability via layer decomposition.arXiv preprint arXiv:2512.15603, 2025

    Yin, S., Zhang, Z., Tang, Z., Gao, K., Xu, X., Yan, K., Li, J., Chen, Y., Chen, Y., Shum, H.Y., et al.: Qwen-image-layered: Towards inherent editability via layer decomposition. arXiv preprint arXiv:2512.15603 (2025) 4

    work page arXiv 2025

  71. [78]

    In: ACM Transactions on Graphics (2023) 2, 3, 5

    Zhang, B., Tang, J., Niessner, M., Wonka, P.: 3DShape2VecSet: A 3D shape repre- sentation for neural fields and generative diffusion models. In: ACM Transactions on Graphics (2023) 2, 3, 5

    work page 2023

  72. [79]

    ACM Transactions On Graphics (TOG)42(4), 1–16 (2023) 10

    Zhang, B., Tang, J., Niessner, M., Wonka, P.: 3dshape2vecset: A 3d shape repre- sentation for neural fields and generative diffusion models. ACM Transactions On Graphics (TOG)42(4), 1–16 (2023) 10

    work page 2023

  73. [80]

    Gs-lrm: Large reconstruction model for 3d gaussian splatting.ArXiv, abs/2404.19702, 2024

    Zhang, K., Bi, S., Tan, H., Xiangli, Y., Zhao, N., Sunkavalli, K., Xu, Z.: GS-LRM: large reconstruction model for 3D Gaussian splatting. arXiv2404.19702(2024) 11

    work page arXiv 2024

  74. [81]

    In: Proc

    Zhang, L., Wang, Z., Zhang, Q., Qiu, Q., Pang, A., Jiang, H., Yang, W., Xu, L., Yu, J.: CLAY: A controllable large-scale generative model for creating high-quality 3D assets. In: Proc. SIGGRAPH (2024) 2, 3, 4, 12, 14, 6, 8, 10, 11

    work page 2024

  75. [82]

    In: Proc

    Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proc. CVPR (2023) 7

    work page 2023

  76. [83]

    arXiv 2503.09439(2025) 11

    Zhang,Q.,Jian,X.,Zhang,X.,Wang,W.,Hou,J.:Supercarver:Texture-consistent 3d geometry super-resolution for high-fidelity surface detail generation. arXiv 2503.09439(2025) 11

    work page arXiv 2025

  77. [84]

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

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 586–595 (2018) 12

    work page 2018

  78. [86]

    Hunyuan3D 2.0: Scaling Diffusion Models for High Resolution Textured 3D Assets Generation

    Zhao, Z., Lai, Z., Lin, Q., Zhao, Y., Liu, H., Yang, S., Feng, Y., Yang, M., Zhang, S., Yang, X., et al.: Hunyuan3d 2.0: Scaling diffusion models for high resolution textured 3d assets generation. arXiv preprint arXiv:2501.12202 (2025) 4

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  79. [87]

    In: Proc

    Zhao, Z., Liu, W., Chen, X., Zeng, X., Wang, R., Cheng, P., Fu, B., Chen, T., Yu, G., Gao, S.: Michelangelo: Conditional 3D shape generation based on shape- image-text aligned latent representation. In: Proc. NeurIPS (2023) 4 3D-ReGen: A Unified 3D Geometry Regeneration Framework 21

    work page 2023

  80. [88]

    VAEEncoder Conditional Image Grid DINO V2 … Self attnCross attn⟳× 𝑡∗… MLPDiffusion Perturbation𝑧!∗=𝛼!∗𝑧

    Zhuang, J., Kang, D., Cao, Y.P., Li, G., Lin, L., Shan, Y.: Tip-editor: An accurate 3d editor following both text-prompts and image-prompts. ACM Transactions on Graphics (TOG)43(4), 1–12 (2024) 13, 11 3D-ReGen: A Unified 3D Geometry Regeneration Framework 1 Supplementary Material of 3D-ReGen: A Unified 3D Geometry Regeneration Framework The supplementary ...

    work page 2024