Pith. sign in

REVIEW 3 major objections 6 minor 108 references

3D editing from a rough box and one image, no precise masks required

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · glm-5.2

2026-07-09 17:47 UTC pith:5QCS3GRT

load-bearing objection Solid end-to-end 3D editing system with a real evaluation gap on the replace task the 3 major comments →

arxiv 2607.07187 v1 pith:5QCS3GRT submitted 2026-07-08 cs.CV

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

Youtan Yin , Yanning Zhou , Jiacheng Wei , Xiaofeng Yang , Jun Zhang , Jiayang Bai , Jingwen Ye , Weidong Zhang
show 1 more author
Guosheng Lin
This is my paper
classification cs.CV
keywords editingeditverse3dobjectcoarseeditedhigh-qualitylossmasks
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper proposes EditVerse3D, an end-to-end framework for local 3D object editing that takes three inputs: the original 3D object, a coarse 3D bounding box drawn loosely around the region to modify, and a single 2D reference image showing the desired change. The framework directly outputs an edited 3D object without requiring pre-edited 2D views, precise 3D masks, or multi-stage 2D-to-3D lifting pipelines. The central technical contributions are a region-aware adaptive loss that reweights training signal to balance learning between the target editing region and the preserved region while emphasizing hard examples, a joint normalization scheme that keeps the input object and its mask spatially aligned, and data augmentation strategies including training with randomly perturbed coarse bounding boxes and filtering low-quality editing pairs. The authors also construct a large-scale training dataset of roughly 85,000 meshes and 500,000 editing pairs by treating part removal and restoration as editing operations, leveraging 3D segmentation data. Experiments show the method outperforms existing approaches on both addition and replacement editing tasks across geometry and texture quality metrics while maintaining inference speed comparable to the underlying 3D generative backbone.

Core claim

The key finding is that a 3D generative model can be adapted into a direct 3D editor that accepts only a coarse bounding box and a single 2D image as guidance, producing high-fidelity local edits, provided three conditions are met: the training loss is explicitly balanced between edited and preserved regions with hard-example emphasis, the input object and mask are jointly normalized to preserve spatial alignment, and the model is trained on augmented coarse masks rather than exact shapes. The authors demonstrate that training with exact 3D masks generalizes poorly to coarse inputs at test time, while training with perturbed bounding boxes closes the train-test gap. They further show that a

What carries the argument

The central mechanism is the region-aware adaptive loss (Equation 6), which decomposes the flow-matching velocity prediction error into three terms: a mean loss over the masked editing region, a mean loss over the unmasked preserved region, and a hard-example loss over the top-τ percent of per-element errors. These terms are balanced by ratios of their magnitudes so that neither the large preserved region nor the small editing region dominates training. This loss operates within a rectified flow model built on the TRELLIS 3D generative backbone, where structure and texture are encoded into separate latent spaces and the flow model learns to predict velocity fields that transform noise into编辑

Load-bearing premise

The framework assumes that training exclusively on 'add' editing pairs—where a removed part is restored to an otherwise intact object—provides sufficient signal for the model to generalize to 'replace' edits where existing geometry is overwritten, a claim validated on a small test set of 100 meshes without ground-truth comparisons for the edited regions.

What would settle it

If a model trained on the authors' 'add'-only dataset fails to produce coherent edits when the target region already contains geometry that must be replaced rather than filled in, the central practical claim—that a single training paradigm covers both addition and replacement editing—would not hold.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the approach generalizes as claimed, 3D content creation tools could allow non-expert users to edit specific parts of 3D objects by drawing a loose box and providing a reference image, removing the need for 3D mask painting or 2D image editing skills.
  • The dataset construction methodology—treating part removal and restoration as editing pairs—could be applied to other 3D generation backbones to create supervised editing training data at scale, potentially accelerating development of editing-capable 3D models beyond this specific framework.
  • The region-aware loss formulation is architecture-agnostic in principle and could transfer to other flow-matching or diffusion-based editing tasks where an imbalance exists between target and preserved regions, including 2D image inpainting or video editing.
  • The finding that training with coarse masks generalizes better than training with exact masks suggests a broader principle: matching training-time input noise to expected inference-time input imprecision may be more important than providing clean supervision signal.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The dataset's reliance on 'add' operations (part restoration) as the primary training signal means the model may learn a restoration-like prior rather than a general editing prior; the demonstrated generalization to 'replace' edits on a small test set without ground truth may not fully validate the broader claim of general-purpose editing capability.
  • The joint normalization requirement implies that the framework assumes the input 3D object and the coarse bounding box share a common coordinate frame at inference time; in practical deployment scenarios where users draw boxes in different interface contexts, this alignment may require additional handling not discussed in the paper.
  • The separate training of structure and texture flow models means that geometric and appearance edits are decoupled; this could produce subtle inconsistencies at boundaries where structure and texture transitions interact, a limitation not explicitly addressed in the evaluation.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. EditVerse3D proposes an end-to-end framework for local 3D object editing that takes a 3D object, a coarse 3D bounding box, and a 2D reference image as input, producing an edited 3D object without requiring precise masks or pre-edited 2D views. The method builds on the TRELLIS 3D generative backbone, introducing a region-aware adaptive loss that balances gradients between edited and preserved regions while emphasizing hard examples, along with data augmentation strategies (joint normalization, coarse mask training, filtering). The authors construct a large-scale dataset of ~85k meshes and ~500k editing pairs by treating part removal and restoration as 'add' editing operations. Experiments compare against Instant3dit, Repaint, FlowEdit, and VoxHammer on addition and replacement tasks.

Significance. The paper addresses a practical gap in 3D editing by relaxing input requirements to coarse bounding boxes, which is a meaningful usability improvement over prior work requiring precise masks or pre-edited 2D views. The region-aware loss formulation (Eqs. 4–6) is a reasonable adaptation of hard-example mining to the editing setting. The large-scale dataset construction from part segmentation data is a useful contribution. Quantitative results on the Add task (Tables 1–2) are comprehensive across multiple metrics and show clear improvements. The ablation study (Table 3) properly isolates individual components.

major comments (3)
  1. §4.1, Table 1 (Replace columns): The paper's central claim asserts 'superior visual quality and quantitative performance' for 3D editing broadly. However, for the Replace task, the authors explicitly state that ground-truth edits are unavailable, so quantitative metrics are computed 'only on non-edited regions.' Table 1's Replace columns therefore measure preservation of unedited areas, not the quality of the replacement itself. Table 2 provides a Target Editing Region vs. Unedited Region breakdown for the Add task but no equivalent breakdown exists for Replace. Consequently, quantitative superiority on the edited content—the more practically important operation—is not substantiated. The claim of quantitative superiority for replacement editing should be either softened to reflect that only preservation is measured, or supplemented with a metric that assesses edit quality (e.g., text- or
  2. §3.4, §4.1: The model is trained exclusively on 'add' editing pairs (part removal/restoration). Generalization to 'replace' edits is validated only on the VoxHammer dataset (100 meshes, 300 pairs) without ground-truth edits for replaced regions. The paper states 'models trained on it can generalize effectively to replace edits,' but the quantitative evidence for this is limited to unedited-region preservation (Table 1, Replace columns). The practical utility of the framework hinges on this generalization, yet it is the weakest-supported claim. A user study or a replace-task benchmark with ground truth would substantially strengthen the contribution.
  3. §3.3, Eq. (6): The overall loss L_edit = L_m + (|L_m|/|L_m̄|)·L_m̄ + (|L_m|/|L_hard|)·L_hard uses the ratio of loss magnitudes as balancing weights. This creates a coupling where the weight on L_m̄ depends on the current value of L_m, which changes during training. The authors do not discuss whether this dynamic weighting is stable or whether it could cause oscillations. Additionally, the hyperparameter τ for hard-example selection (Eq. 5) is not reported in the main text. These details affect reproducibility.
minor comments (6)
  1. §3.1: The forward process is defined as x_t = (1−t)x_0 + tε, but the standard rectified flow convention (Lipman et al. [47]) typically uses x_t = (1−t)x_0 + tε or x_t = t·x_0 + (1−t)·ε depending on the direction convention. Clarifying which convention is used would help readers.
  2. Table 3 caption: 'The CD metric is scaled by 10² for better visualization' is noted, but the CD values in Table 1 are not scaled, making cross-table comparison of CD values confusing. A consistent scaling convention would help.
  3. §4.1: The test set for 'add' is described as 'about 200 meshes and 1500 3D editing pairs from PartObjaverse-Tiny.' Exact numbers would improve reproducibility.
  4. Fig. 2: The architecture diagram is informative but the text labels are small. Consider enlarging key components for readability.
  5. §3.3: The volume threshold for filtering 'unrealistic editing pairs' is mentioned but its specific value is not provided in the main text (the authors note it is in supplementary materials). Including it here would be helpful.
  6. References: Several arXiv preprints are cited with future dates (e.g., [3] Dec 2024, [10] Jul 2025, [11] Jul 2024). Verify these dates are correct.

Circularity Check

0 steps flagged

No circularity found: derivation chain is self-contained with empirical validation against external benchmarks

full rationale

The paper's core technical contributions—a region-aware loss reweighting (Eq. 4-6), joint normalization, coarse mask augmentation, and dataset construction—are all standard techniques (hard-example mining, loss balancing, data augmentation) applied in a new context. None of these reduce to their inputs by construction. The loss formulation decomposes standard MSE into masked/non-masked regions and adds a hard-example term; this is a training strategy, not a derivation that assumes its conclusion. The TRELLIS backbone [86] is cited from external authors (Xiang et al.). The Partverse dataset [70] is also externally sourced. No load-bearing self-citations exist in the derivation chain. The central claim of quantitative superiority is evaluated against external baselines (Instant3dit, Repaint, FlowEdit, VoxHammer) on independently constructed test sets (PartObjaverse-Tiny, VoxHammer dataset). The ablation study (Table 3) uses proper controls with independent variables. The skeptic's concern about Replace-task metrics being computed only on preserved regions (Sec. 4.1) is a legitimate evaluation completeness issue, but it is not circularity—the paper is transparent about this limitation and does not attempt to disguise it as a derivation or prediction. The generalization claim from 'add' to 'replace' edits is empirically tested, not derived from a self-referential argument. No step in the paper's chain reduces to its own inputs by definition, fit, or self-citation.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The paper introduces several free parameters whose values are not fully specified in the main text (tau, perturbation ranges, filtering threshold), which affects reproducibility. The core axioms are domain assumptions about the TRELLIS backbone and the part-removal-as-editing proxy. The invented entities (loss function, dataset) are methodological contributions validated through experiments rather than physical postulates.

free parameters (4)
  • tau (hard-example percentage) = Not explicitly stated in the paper
    The top-tau% threshold for selecting hard examples in the region-aware loss (Eq. 5) is a free parameter that determines which regions are emphasized during training. Its value is not specified in the main text.
  • Loss balancing weights = Derived from loss magnitudes (Eq. 6)
    The weights |Lm|/|L_bar_m| and |Lm|/|Lhard| in the overall loss (Eq. 6) are computed from current loss magnitudes, introducing an adaptive coupling whose stability depends on the relative scale of the three loss terms.
  • Bounding box perturbation parameters = Not explicitly stated
    Random disturbances applied to the bounding box size and position during training (Sec. 3.3, 'Training with Coarse 3D Masks') are described qualitatively but the specific scale and translation ranges are not provided.
  • Volume threshold for filtering = Not explicitly stated
    The threshold below which target editing regions are filtered out as 'too small' (Sec. 3.3) is a free parameter whose value is deferred to supplementary materials.
axioms (4)
  • domain assumption TRELLIS encoder-decoder provides nearly lossless perceptual compression of 3D objects
    The entire editing framework is built on the assumption that TRELLIS latents faithfully represent 3D objects. Stated in Sec. 3.1: 'TRELLIS's encoder-decoder network is highly effective at compressing and restoring 3D objects with nearly lossless perceptual quality.'
  • ad hoc to paper Part removal and restoration serves as a valid proxy for general 3D editing operations
    The dataset construction (Sec. 3.4) assumes that treating part removal/restoration as 'add' editing provides training signal that generalizes to 'replace' and other edit types. The paper provides limited empirical support for this on the VoxHammer dataset.
  • domain assumption Rectified flow models trained on editing pairs can generalize to novel objects and edits at inference
    Standard assumption in generative model training; the model learns a conditional distribution over edits that applies to unseen inputs. Validated empirically through test set evaluation.
  • domain assumption Joint normalization of 3D object and mask preserves spatial relationships necessary for editing
    Stated in Sec. 3.3 and illustrated in Fig. 3. The paper notes that separate normalization distorts relative positions and prevents convergence, but this is presented as an empirical observation rather than a proven property.
invented entities (2)
  • Region-aware adaptive loss (L_edit) independent evidence
    purpose: Balances training between masked and non-masked regions while emphasizing hard-to-learn areas
    The loss function is validated through ablation (Table 3, Exp. #4 vs #5) showing improved performance over standard MSE loss. It is a training strategy, not a physical entity.
  • 3D editing dataset from parts information independent evidence
    purpose: Provides supervised training pairs for 3D editing by treating part removal/restoration as editing operations
    The dataset is constructed from existing 3D repositories (Partverse, Objaverse) with a described pipeline. Its utility is validated through model training and evaluation.

pith-pipeline@v1.1.0-glm · 22892 in / 3129 out tokens · 413294 ms · 2026-07-09T17:47:36.580395+00:00 · methodology

0 comments
read the original abstract

Local editing of 3D objects remains a long-standing challenge. When interacting with 3D content, humans naturally tend to specify a coarse region of interest for modification rather than defining precise editing boundaries. However, previous methods rely on fully edited 2D images, precise 3D masks, or redundant pipelines, which present a gap. To bridge this gap, we propose EditVerse3D, a novel 3D editing framework that enables high-quality object editing under such coarse guidance. Our approach takes as input a 3D object to be edited, a coarse 3D bounding box indicating the target region, and a reference 2D image describing the desired modification. It produces a coherent, high-fidelity edited 3D object. To facilitate this editing, we introduce a novel region-aware adaptive loss that emphasizes hard-to-learn regions and balances the objective between target and preserved areas. Complementing our loss function, we enhance model robustness and generalization through targeted data augmentations, such as training with scaled 3D masks and filtering out unrealistic editing pairs. We construct a large-scale 3D editing dataset derived from parts information. Extensive experiments demonstrate that EditVerse3D achieves superior visual quality and quantitative performance compared to existing 3D editing approaches. Please visit our project page at https://editverse3d.github.io.

Figures

Figures reproduced from arXiv: 2607.07187 by Guosheng Lin, Jiacheng Wei, Jiayang Bai, Jingwen Ye, Jun Zhang, Weidong Zhang, Xiaofeng Yang, Yanning Zhou, Youtan Yin.

Figure 1
Figure 1. Figure 1: Editing results of our method. Given a 3D object, a user-specified coarse 3D bounding box indicating the target editing region, and an image prompt defining the editing goal, our approach generates high-quality, coherent edits. Our method does not require fully edited 2D views, precise 3D masks, or redundant pipelines. the back-and-forth between 3D and 2D introduces cumulative errors, resulting in a declin… view at source ↗
Figure 2
Figure 2. Figure 2: An overview of our method. Given a masked 3D object as input, we first extract its structure and texture latents using the TRELLIS encoder. The input latents are concatenated with a binary mask and random noise along the feature channel dimension, then fed into the flow-matching model, which takes the editing target as a condition. The flow model generates edited latents, which decode into the final result… view at source ↗
Figure 3
Figure 3. Figure 3: Effect of joint normalization. Separate normalization leads to overlapping (3), whereas our joint normalization preserves the correct spatial relationship (6). non-masked regions contribute equally to the overall loss, even if one region is significantly larger than the other. Suppose M is the boolean mask, and m and m¯ are the number of elements in the masked and nonmasked region, respectively. The loss t… view at source ↗
Figure 4
Figure 4. Figure 4: Overview of our 3D editing dataset. Left: Parts Distribution across the dataset. Right: An Example Editing Pair illustrating the ground truth (top left), seg￾mentation (top right), source object (bottom left), and editing target (bottom right). Training with Coarse 3D Masks. In practical scenarios, users prefer to provide a coarse 3D bounding box that is larger than the actual region to be edited rather th… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison. Our method yields prompt-aligned edits while pre￾serving unedited regions. Repaint aligns structures in rows (1, 4) but misaligns textures in (2, 3). VoxHammer shows partial consistency in (1, 2) but fails in (3, 4). Qualitative. We visualize the editing results produced by our method and existing approaches in [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative results. See additional examples in the supplementary material [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative ablation results. 3D Object displays the masked object alongside the target prompt (a sword). Exact Mask (Exp.#1) is trained with precise 3D masks. Ours BBox+ (Exp.#4) utilizes our mask augmentation strategy. Ours Loss (Exp.#5) further incorporates our region-aware loss. achieves good results on the training set, it generalizes poorly to the test set and produces meaningless outputs. Filtering … view at source ↗
Figure 8
Figure 8. Figure 8: Robustness to input variations. Base shows editing results using an ideal bounding box and a 2D prompt. Scale enlarges the baseline box by 1.5×. Shift trans￾lates the scaled box to the right. Finally, Pose demonstrates results when the image prompt (Cond.) features a different viewing angle than the Base [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗

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

108 extracted references · 108 canonical work pages · 39 internal anchors

  1. [1]

    Bar-On, R., Cohen-Bar, D., Cohen-Or, D.: Editp23: 3d editing via propagation of image prompts to multi-view (2025),https://arxiv.org/abs/2506.20652

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  2. [2]

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

    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 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16273–16282 (2025)

    work page 2025

  3. [3]

    SceneFactor: Factored Latent 3D Diffusion for Controllable 3D Scene Generation

    Bokhovkin, A., Meng, Q., Tulsiani, S., Dai, A.: Scenefactor: Factored latent 3d diffusion for controllable 3d scene generation (Dec 2024).https://doi.org/10. 48550/arXiv.2412.01801

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  4. [4]

    Cen, J., Fang, J., Yang, C., Xie, L., Zhang, X., Shen, W., Tian, Q.: Segment any 3d gaussians (Feb 2025).https://doi.org/10.48550/arXiv.2312.00860

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2312.00860 2025

  5. [5]

    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 (2024),https: //arxiv.org/abs/2403.12032

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  6. [6]

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

    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/CVF Conference on Computer Vision and Pattern Recognition. pp. 5881–5892 (2025)

    work page 2025

  7. [7]

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

    Chen, M., Xie, J., Laina, I., Vedaldi, A.: Shap-editor: Instruction-guided latent 3d editing in seconds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26456–26466 (2024)

    work page 2024

  8. [8]

    Chen, T., Yu, C., Li, J., Zhang, J., Zhu, L., Ji, D., Zhang, Y., Zang, Y., Li, Z., Sun, L.: Reasoning3d – grounding and reasoning in 3d: Fine-grained zero-shot open-vocabulary 3d reasoning part segmentation via large vision-language models (May 2024).https://doi.org/10.48550/arXiv.2405.19326

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2405.19326 2024

  9. [9]

    Plasticine3D: 3D Non-Rigid Editing with Text Guidance by Multi-View Embedding Optimization

    Chen, Y., Hu, T., Tang, Y., Chen, S., Chen, A., Yi, R.: Plasticine3d: 3d non-rigid editing with text guidance by multi-view embedding optimization (Jul 2024). https://doi.org/10.48550/arXiv.2312.10111 16 Y. Yin et al

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2312.10111 2024

  10. [10]

    Chen,Y.,Li,Z.,Wang,Y.,Zhang,H.,Li,Q.,Zhang,C.,Lin,G.:Ultra3d:Efficient and high-fidelity 3d generation with part attention (Jul 2025).https://doi.org/ 10.48550/arXiv.2507.17745

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2507.17745 2025

  11. [11]

    ACM Transactions on Graphics (TOG) (Jul 2024)

    ChoiChangwoon, LeeJaeah, ParkJaesik, Min, K.: 3doodle: Compact abstraction of objects with 3d strokes. ACM Transactions on Graphics (TOG) (Jul 2024). https://doi.org/10.1145/3658156

    work page doi:10.1145/3658156 2024

  12. [12]

    Darcet, T., Oquab, M., Mairal, J., Bojanowski, P.: Vision transformers need reg- isters (2023)

    work page 2023

  13. [13]

    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. arXiv preprint arXiv:2307.05663 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  14. [14]

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

    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: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13142–13153 (2023)

    work page 2023

  15. [15]

    In: Forty-Second International Conference on Machine Learning (Jun 2025)

    Deng, Y., He, X., Mei, C., Wang, P., Tang, F.: Fireflow: Fast inversion of rectified flow for image semantic editing. In: Forty-Second International Conference on Machine Learning (Jun 2025)

    work page 2025

  16. [16]

    Dinh, N.A., Lang, I., Kim, H., Stein, O., Hanocka, R.: Geometry in style: 3d styl- ization via surface normal deformation (2025),https://arxiv.org/abs/2503. 23241

    work page 2025

  17. [17]

    Dong, S., Ding, L., Huang, Z., Wang, Z., Xue, T., Xu, D.: Interactive3d: Create what you want by interactive 3d generation (2024),https://arxiv.org/abs/ 2404.16510

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  18. [18]

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

    Erkoç, Z., Gümeli, C., Wang, C., Nießner, M., Dai, A., Wonka, P., Lee, H.Y., Zhuang, P.: Preditor3d: Fast and precise 3d shape editing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 640–649 (2025)

    work page 2025

  19. [19]

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

    Fan, H., Su, H., Guibas, L.J.: A point set generation network for 3d object re- construction from a single image. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 605–613 (2017)

    work page 2017

  20. [20]

    In: Leonardis, A., Ricci, E., Roth, S., Russakovsky, O., Sattler, T., Varol, G

    Fang, S., Wang, Y., Tsai, Y.H., Yang, Y., Ding, W., Zhou, S., Yang, M.H.: Chat- edit-3d: Interactive 3d scene editing via text prompts. In: Leonardis, A., Ricci, E., Roth, S., Russakovsky, O., Sattler, T., Varol, G. (eds.) Computer Vision – ECCV 2024. pp. 199–216. Springer Nature Switzerland, Cham (2025).https: //doi.org/10.1007/978-3-031-72946-1_12

    work page doi:10.1007/978-3-031-72946-1_12 2024

  21. [21]

    org/10.48550/arXiv.2411.19322

    Fischer, M., Georgiev, I., Groueix, T., Kim, V.G., Ritschel, T., Deschaintre, V.: Sama: Material-aware 3d selection and segmentation (Nov 2024).https://doi. org/10.48550/arXiv.2411.19322

    work page doi:10.48550/arxiv.2411.19322 2024

  22. [22]

    Gao, W., Wang, D., Fan, Y., Bozic, A., Stuyck, T., Li, Z., Dong, Z., Ranjan, R., Sarafianos, N.: 3d mesh editing using masked lrms (2025),https://arxiv.org/ abs/2412.08641

    work page arXiv 2025

  23. [23]

    Gao, Z., Yi, R., Huang, Y., Chen, W., Zhu, C., Xu, K.: Self-supervised learning of hybrid part-aware 3d representations of 2d gaussians and superquadrics (Jul 2025).https://doi.org/10.48550/arXiv.2408.10789

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2408.10789 2025

  24. [24]

    Google: Google gemini.https://gemini.google.com/app(2025)

    work page 2025

  25. [25]

    ARAP-GS: Drag-driven As-Rigid-As-Possible 3D Gaussian Splatting Editing with Diffusion Prior

    Han, X., Tian, R., Tong, Y., Yu, F., Liu, D., Zhang, Y.: Arap-gs: Drag-driven as-rigid-as-possible 3d gaussian splatting editing with diffusion prior (Apr 2025). https://doi.org/10.48550/arXiv.2504.12788 EditVerse3D: High-Quality 3D Object Editing with Region-Aware Learning 17

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2504.12788 2025

  26. [26]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    He, K., Wu, C.H., Gilitschenski, I.: Ctrl-d: Controllable dynamic 3d scene edit- ing with personalized 2d diffusion. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 26630–26640 (2025)

    work page 2025

  27. [27]

    He, X., Zou, Z.X., Chen, C.H., Guo, Y.C., Liang, D., Yuan, C., Ouyang, W., Cao, Y.P., Li, Y.: Sparseflex: High-resolution and arbitrary-topology 3d shape modeling (Mar 2025).https://doi.org/10.48550/arXiv.2503.21732

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2503.21732 2025

  28. [28]

    In: Advances in Neural Information Processing Systems

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. In: Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc. (2017)

    work page 2017

  29. [29]

    Towards Generalized and Training-Free Text-Guided Semantic Manipulation

    Hong, Y., Cai, X., Zeng, P., Zhang, S., Song, J., Gao, L., Shen, H.T.: To- wards generalizedand training-freetext-guided semanticmanipulation (Jul2025). https://doi.org/10.48550/arXiv.2504.17269

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2504.17269

  30. [30]

    https://doi.org/10.48550/arXiv

    Jincheng, J., Cai, Y., Liu, L.: Craftmesh: High-fidelity generative mesh manipula- tion via poisson seamless fusion (Sep 2025).https://doi.org/10.48550/arXiv. 2509.13688

    work page internal anchor Pith review doi:10.48550/arxiv 2025

  31. [31]

    Jose, C., Moutakanni, T., Kang, D., Baldassarre, F., Darcet, T., Xu, H., Li, D., Szafraniec, M., Ramamonjisoa, M., Oquab, M., Siméoni, O., Vo, H.V., Labatut, P., Bojanowski, P.: Dinov2 meets text: A unified framework for image- and pixel- level vision-language alignment (2024)

    work page 2024

  32. [32]

    In: Leonardis, A., Ricci, E., Roth, S., Russakovsky, O., Sattler, T., Varol, G

    Ju, X., Liu, X., Wang, X., Bian, Y., Shan, Y., Xu, Q.: Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. In: Leonardis, A., Ricci, E., Roth, S., Russakovsky, O., Sattler, T., Varol, G. (eds.) Computer Vision – ECCV 2024. pp. 150–168. Springer Nature Switzerland, Cham (2025). https://doi.org/10.1007/978-3-031-72661-3_9

    work page doi:10.1007/978-3-031-72661-3_9 2024

  33. [33]

    ACM Transactions on Graphics42(4), 1–14 (Aug 2023).https://doi.org/10.1145/3592433

    Kerbl, B., Kopanas, G., Leimkuehler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics42(4), 1–14 (Aug 2023).https://doi.org/10.1145/3592433

    work page doi:10.1145/3592433 2023

  34. [34]

    In: The Thirteenth International Conference on Learning Representations (Oct 2024)

    Kim, J., Lee, S., Shin, J., Choi, J., Shim, H.: Dreamcatalyst: Fast and high-quality 3d editing via controlling editability and identity preservation. In: The Thirteenth International Conference on Learning Representations (Oct 2024)

    work page 2024

  35. [35]

    Kulikov, V., Kleiner, M., Huberman-Spiegelglas, I., Michaeli, T.: Flowedit: Inversion-free text-based editing using pre-trained flow models (Jul 2025).https: //doi.org/10.48550/arXiv.2412.08629

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2412.08629 2025

  36. [36]

    Labs, B.F., Batifol, S., Blattmann, A., Boesel, F., Consul, S., Diagne, C., Dock- horn, 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 (2025),https://a...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  37. [37]

    Lai, Z., Zhao, Y., Zhao, Z., Liu, H., Wang, F., Shi, H., Yang, X., Lin, Q., Huang, J., Liu, Y., Jiang, J., Guo, C., Yue, X.: Unleashing vecset diffusion model for fast shape generation (Mar 2025).https://doi.org/10.48550/arXiv.2503.16302

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2503.16302 2025

  38. [38]

    Inter- national Conference on Representation Learning2025, 97419–97443 (May 2025)

    Lan, Y., Zhou, S., Lyu, Z., Hong, F., Yang, S., Dai, B., Pan, X., Loy, C.C.: Gaussiananything: Interactive point cloud flow matching for 3d generation. Inter- national Conference on Representation Learning2025, 97419–97443 (May 2025)

    work page 2025

  39. [39]

    Le, D.H., Pham, T., Kembhavi, A., Mandt, S., Ma, W.C., Lu, J.: Preserving identity with variational score for general-purpose 3d editing (Jun 2024).https: //doi.org/10.48550/arXiv.2406.08953

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2406.08953 2024

  40. [40]

    05929 18 Y

    Li, H., Tian, Y., Wang, Y., Liao, Y., Wang, L., Wang, Y., Zhou, P.Y.: Text-to-3d generation by 2d editing (Mar 2025).https://doi.org/10.48550/arXiv.2412. 05929 18 Y. Yin et al

    work page doi:10.48550/arxiv.2412 2025

  41. [41]

    Li, H., Erkoc, Z., Li, L., Sirigatti, D., Rozov, V., Dai, A., Nießner, M.: Mesh- pad: Interactive sketch-conditioned artist-reminiscent mesh generation and edit- ing (Aug 2025).https://doi.org/10.48550/arXiv.2503.01425

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2503.01425 2025

  42. [42]

    VoxHammer: Training-Free Precise and Coherent 3D Editing in Native 3D Space

    Li, L., Huang, Z., Feng, H., Zhuang, G., Chen, R., Guo, C., Sheng, L.: Voxham- mer: Training-free precise and coherent 3d editing in native 3d space (Aug 2025). https://doi.org/10.48550/arXiv.2508.19247

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2508.19247 2025

  43. [44]

    Li, W., Zhang, X., Sun, Z., Qi, D., Li, H., Cheng, W., Cai, W., Wu, S., Liu, J., Wang, Z., Chen, X., Tian, F., Pan, J., Li, Z., Yu, G., Zhang, X., Jiang, D., Tan, P.: Step1x-3d: Towards high-fidelity and controllable generation of textured 3d assets (May 2025).https://doi.org/10.48550/arXiv.2505.07747

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2505.07747 2025

  44. [45]

    Lin, J., Yang, X., Chen, M., Xu, Y., Yan, D., Wu, L., Xu, X., Xu, L., Zhang, S., Chen, Y.C.: Kiss3dgen: Repurposing image diffusion models for 3d asset genera- tion.In:ProceedingsoftheComputerVisionandPatternRecognitionConference. pp. 5870–5880 (2025)

    work page 2025

  45. [46]

    Lin, Y., Lin, C., Pan, P., Yan, H., Feng, Y., Mu, Y., Fragkiadaki, K.: Partcrafter: Structured 3d mesh generation via compositional latent diffusion transformers (Jun 2025).https://doi.org/10.48550/arXiv.2506.05573

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2506.05573 2025

  46. [47]

    Lipman, Y., Chen, R.T.Q., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling (2023),https://arxiv.org/abs/2210.02747

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  47. [48]

    Liu, F., Wang, H., Chen, W., Sun, H., Duan, Y.: Make-your-3d: Fast and consis- tent subject-driven 3d content generation (2024),https://arxiv.org/abs/2403. 09625

    work page 2024

  48. [49]

    ACM Transactions on Graphics43(4), 1–13 (Jul 2024)

    Liu, F.L., Fu, H., Lai, Y.K., Gao, L.: Sketchdream: Sketch-based text-to-3d gen- eration and editing. ACM Transactions on Graphics43(4), 1–13 (Jul 2024). https://doi.org/10.1145/3658120

    work page doi:10.1145/3658120 2024

  49. [50]

    Liu, M., Uy, M.A., Xiang, D., Su, H., Fidler, S., Sharp, N., Gao, J.: Partfield: Learning 3d feature fields for part segmentation and beyond (2025)

    work page 2025

  50. [51]

    In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition

    Long, X., Guo, Y.C., Lin, C., Liu, Y., Dou, Z., Liu, L., Ma, Y., Zhang, S.H., Habermann, M., Theobalt, C., Wang, W.: Wonder3d: Single image to 3d using cross-domain diffusion. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 9970–9980 (2024)

    work page 2024

  51. [52]

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

    Lu, R., Chen, Y., Ni, J., Jia, B., Liu, Y., Wan, D., Zeng, G., Huang, S.: Movis: Enhancing multi-object novel view synthesis for indoor scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26767–26778 (2025)

    work page 2025

  52. [53]

    Lu, Y., Tian, Y., Jiang, Z., Zhao, Y., Yang, Y., Ouyang, H., Hu, H., Yu, H., Shen, Y., Liao, Y.: Orientation matters: Making 3d generative models orientation- aligned (Jun 2025).https://doi.org/10.48550/arXiv.2506.08640

    work page doi:10.48550/arxiv.2506.08640 2025

  53. [54]

    In: 2022 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR)

    Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L.: Repaint: Inpainting using denoising diffusion probabilistic models. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11451–11461 (Jun 2022).https://doi.org/10.1109/CVPR52688.2022.01117

    work page doi:10.1109/cvpr52688.2022.01117 2022

  54. [55]

    arXiv preprint arXiv:2509.06784 (2025) EditVerse3D: High-Quality 3D Object Editing with Region-Aware Learning 19

    Ma, C., Li, Y., Yan, X., Xu, J., Yang, Y., Wang, C., Zhao, Z., Guo, Y., Chen, Z., Guo, C.: P3-sam: Native 3d part segmentation. arXiv preprint arXiv:2509.06784 (2025) EditVerse3D: High-Quality 3D Object Editing with Region-Aware Learning 19

    work page arXiv 2025

  55. [56]

    Find Any Part in 3D

    Ma, Z., Yue, Y., Gkioxari, G.: Find any part in 3d (Mar 2025).https://doi. org/10.48550/arXiv.2411.13550

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2411.13550 2025

  56. [57]

    In: International Conference on Learning Representations (Oct 2021)

    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: International Conference on Learning Representations (Oct 2021)

    work page 2021

  57. [58]

    In: Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision

    Mikaeili, A., Perel, O., Safaee, M., Cohen-Or, D., Mahdavi-Amiri, A.: Sked: Sketch-guided text-based 3d editing. In: Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision. pp. 14607–14619 (2023)

    work page 2023

  58. [59]

    Com- munications of the ACM65(1), 99–106 (Jan 2022).https://doi.org/10.1145/ 3503250

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Com- munications of the ACM65(1), 99–106 (Jan 2022).https://doi.org/10.1145/ 3503250

    work page 2022

  59. [60]

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Howes, R., Huang, P.Y., Xu, H., Sharma, V., Li, S.W., Galuba, W., Rabbat, M., Assran, M., Ballas, N., Synnaeve, G., Misra, I., Jegou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual feat...

    work page 2023

  60. [61]

    48550/arXiv.2509.00269

    Parelli, M., Oechsle, M., Niemeyer, M., Tombari, F., Geiger, A.: 3d-latte: Latent space 3d editing from textual instructions (Sep 2025).https://doi.org/10. 48550/arXiv.2509.00269

    work page arXiv 2025

  61. [62]

    In: The Eleventh International Conference on Learning Representations (Sep 2022)

    Poole,B.,Jain,A.,Barron,J.T.,Mildenhall,B.:Dreamfusion:Text-to-3dusing2d diffusion. In: The Eleventh International Conference on Learning Representations (Sep 2022)

    work page 2022

  62. [63]

    Qi, Z., Yang, Y., Zhang, M., Xing, L., Wu, X., Wu, T., Lin, D., Liu, X., Wang, J., Zhao, H.: Tailor3d: Customized 3d assets editing and generation with dual-side images (2024),https://arxiv.org/abs/2407.06191

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  63. [64]

    In: Proceed- ings of the Computer Vision and Pattern Recognition Conference

    Ren, Y., Jiang, Z., Zhang, T., Forchhammer, S., Süsstrunk, S.: Fds: Frequency- aware denoising score for text-guided latent diffusion image editing. In: Proceed- ings of the Computer Vision and Pattern Recognition Conference. pp. 2651–2660 (2025)

    work page 2025

  64. [65]

    In: Leonardis, A., Ricci, E., Roth, S., Russakovsky, O., Sattler, T., Varol, G

    Rojas, S., Philip, J., Zhang, K., Bi, S., Luan, F., Ghanem, B., Sunkavalli, K.: Datenerf: Depth-aware text-based editing of nerfs. In: Leonardis, A., Ricci, E., Roth, S., Russakovsky, O., Sattler, T., Varol, G. (eds.) Computer Vision – ECCV

    work page

  65. [66]

    pp. 267–284. Springer Nature Switzerland, Cham (2025).https://doi. org/10.1007/978-3-031-73247-8_16

    work page doi:10.1007/978-3-031-73247-8_16 2025

  66. [67]

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

    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 (CVPR). pp. 10684– 10695 (Jun 2022)

    work page 2022

  67. [68]

    In: The Thirteenth International Conference on Learning Representations (Oct 2024)

    Rout, L., Chen, Y., Ruiz, N., Caramanis, C., Shakkottai, S., Chu, W.S.: Semantic image inversion and editing using rectified stochastic differential equations. In: The Thirteenth International Conference on Learning Representations (Oct 2024)

    work page 2024

  68. [69]

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

    Sella, E., Atia, N., Mokady, R., Averbuch-Elor, H.: Blended point cloud diffu- sion for localized text-guided shape editing. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19119–19129 (2025)

    work page 2025

  69. [70]

    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)

    work page 2023

  70. [71]

    In: ICCV (2025) 20 Y

    Shaocong, D., Lihe, D., Xiao, C., Yaokun, L., Yuxin, W., Yucheng, W., Qi, W., Jaehyeok, K., Chenjian, G., Zhanpeng, H., Zibin, W., Tianfan, X., Xu, D.: From one to more: Contextual part latents for 3d generation. In: ICCV (2025) 20 Y. Yin et al

    work page 2025

  71. [72]

    48550/arXiv.2507.02477

    Song, G., Zhao, Z., Weng, H., Zeng, J., Jia, R., Gao, S.: Mesh silksong: Auto- regressive mesh generation as weaving silk (Jul 2025).https://doi.org/10. 48550/arXiv.2507.02477

    work page arXiv 2025

  72. [73]

    Tang, G., Zhao, W., Ford, L., Benhaim, D., Zhang, P.: Segment any mesh (Mar 2025).https://doi.org/10.48550/arXiv.2408.13679

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2408.13679 2025

  73. [74]

    Tang, J., Lu, R., Li, Z., Hao, Z., Li, X., Wei, F., Song, S., Zeng, G., Liu, M.Y., Lin, T.Y.: Efficient part-level 3d object generation via dual volume packing (Jun 2025).https://doi.org/10.48550/arXiv.2506.09980

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2506.09980 2025

  74. [75]

    Team, T.H.: Hunyuan3d 1.0: A unified framework for text-to-3d and image-to-3d generation (2024)

    work page 2024

  75. [76]

    Team, T.H.: Hunyuan3d 2.0: Scaling diffusion models for high resolution textured 3d assets generation (2025)

    work page 2025

  76. [77]

    Team, T.H.: Hunyuan3d 2.5: Towards high-fidelity 3d assets generation with ul- timate details (2025),https://arxiv.org/abs/2506.16504

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  77. [78]

    Team, T.H.: Hunyuan3d-omni: A unified framework for controllable generation of 3d assets (2025),https://arxiv.org/abs/2509.21245

    work page arXiv 2025

  78. [79]

    In: Proceedings of the 42nd International Conference on Machine Learning

    Wang, J., Pu, J., Qi, Z., Guo, J., Ma, Y., Huang, N., Chen, Y., Li, X., Shan, Y.: Taming rectified flow for inversion and editing. In: Proceedings of the 42nd International Conference on Machine Learning. pp. 64044–64058. PMLR (Oct 2025)

    work page 2025

  79. [80]

    In: Proceedings of the Computer Vision and Pat- tern Recognition Conference

    Wang, R., Xu, S., Dai, C., Xiang, J., Deng, Y., Tong, X., Yang, J.: Moge: Un- locking accurate monocular geometry estimation for open-domain images with optimal training supervision. In: Proceedings of the Computer Vision and Pat- tern Recognition Conference. pp. 5261–5271 (2025)

    work page 2025

  80. [81]

    Image quality assessment: From error visibility to structural similarity.Image Processing, IEEE Transactions on, 13:600 – 612, 05 2004

    Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E.: Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing 13(4), 600–612 (Apr 2004).https://doi.org/10.1109/TIP.2003.819861

    work page doi:10.1109/tip.2003.819861 2004

Showing first 80 references.