Pith. sign in

REVIEW 5 major objections 6 minor 3 cited by

Pro3D-Editor : A Progressive-Views Perspective for Consistent and Precise 3D Editing

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Text-guided 3D edits stay consistent when they start from the view that carries the edit.

desk verdict A genuinely new progressive-views framing in 3D editing, but the evaluation is too weak to back the headline superiority claims. read the letter →

arxiv 2506.00512 v2 pith:MPFQZM2J submitted 2025-05-31 cs.GR cs.AI

Yang Zheng , Mengqi Huang , Nan Chen , Zhendong Mao This is my paper
classification cs.GRcs.AI
keywords text-guided3Deditingprogressive-viewsparadigmediting-salientviewmulti-viewconsistencyGaussianSplattingMixture-of-View-ExpertsLoRAdiffusion-basedsaliencescoring
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that text-guided 3D editing goes wrong when every rendered view is edited with equal weight. Instead, each edit instruction has a most editing-salient view, and consistency comes from propagating that view's edit to the other views, where the edit is sparser. Pro3D-Editor implements this by scoring all rendered views, editing the top-scoring primary view with a 2D editor, transferring the edit to six key views through a direction-aware low-rank adapter in a multi-view diffusion model, and then repairing novel views while updating a 3D Gaussian Splatting object. On six objects and fifteen prompts it reports 47.4% better LPIPS and 9.7% better DINO-I than the compared baselines. If the claim holds, text-guided 3D editing can be both locally precise and multi-view consistent by ordering views by salience instead of treating them equally.

What carries the argument

The central object is the salience ordering plus a directional low-rank adapter. The Primary-view Sampler scores each rendered frame with Eq. (1), $$\mathrm{score}_i = P(I_c,y_s)_i + P(I_c,y_e)_i - \$\alpha$ (P(I_c,y_s,y_e)_p + P(I_c,y_s,y_e)_q),$$ where $P$ is softmaxed CLIP similarity and $p,q$ are the views 135 degrees and 225 degrees from the candidate, with $\alpha = 0.5$; the top frame becomes the primary view. MoVE-LoRA then restructures the multi-view attention of the MV-Adapter backbone: a shared low-rank matrix $A$ is updated only from the primary view, while per-view expert matrices $B_i$ map the primary view's features onto each key view, so the edit flows one way and sparse views cannot overwrite the anchor. A two-stage inference pass blends the backbone's multi-view attention with the fine-tuned one inside editing-region masks to preserve the model's spatial understanding. The Full-view Refiner fine-tunes a ControlNet-Tile-based repair module on edited key views paired with degraded rendered views and uses it to generate repaired novel views that provide dense supervision for the final 3DGS update.

What would settle it

Render a 3D object whose edit target is only visible from a non-frontal azimuth (for example, a stamp on the back), run the Primary-view Sampler with the prompt naming that target, and check whether the highest-scoring frame actually shows the target. Because the scoring rule adds a fixed 0.5 penalty to views 135 degrees and 225 degrees from the candidate, a correct selection for such a prompt would weaken the penalty's assumed geometry, while a wrong selection would show the salience proxy has failed.

Watch

Extended reading notes

Core claim

The paper's central claim is that ideal consistent 3D editing can be achieved through a progressive-views paradigm, in which editing semantics flow from the editing-salient view to editing-sparse views. Concretely, Pro3D-Editor first renders a 360-degree video of the 3D object and scores each frame with softmaxed CLIP similarity to both the object description and the edit text, penalizing views at relative azimuths of 135 and 225 degrees, and picks the top frame as the primary view. A 2D editor edits that frame. Then a multi-view diffusion model fine-tuned with MoVE-LoRA propagates the edit to the other key views: a shared low-rank matrix A is updated only from the primary view, while per-view expert matrices B_i map that primary-view feature to each other view, so correspondences are learned in one direction only. A two-stage inference with editing-region masks preserves the backbone's spatial understanding. Finally, a Full-view Refiner projects the edited key views into 3D, uses a diffusion model to repair newly rendered novel views, and iteratively updates the 3DGS object, giving the reported gains in editing accuracy and spatial consistency.

Load-bearing premise

The whole method stands on one score: a CLIP text-image similarity score, with a fixed penalty for views at 135 degrees and 225 degrees, must correctly pick the view that really shows the requested edit. If it picks the wrong anchor, every later stage inherits the mistake.

Editorial extensions

If this is right

  • Editing instructions that are naturally directional should no longer cause conflicts: front-side edits such as glasses and back-side edits such as a ponytail both propagate from the view that carries them, so the same local region should read consistently from all viewpoints.
  • The two-stage masked inference should let unrelated parts of the object survive the edit, because the second generation round only replaces the masked editing regions with fused attention features.
  • The Full-view Refiner should eliminate the floating Gaussians and structural fragmentation that appear when edited sparse views are projected straight into 3DGS, because it adds repaired novel-view supervision before the final 3D update.
  • The naive fine-tuning baseline with a randomly chosen anchor performs much worse in the paper's ablations, which supports the paper's position that the progressive ordering itself, not the fine-tuning alone, is what delivers consistency.

Reading between the lines

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

  • A natural extension the authors leave implicit is to replace the CLIP-softmax salience rule with a spatial grounding model that localizes the edit text in the rendered views; such a model could handle prompts whose target appears only in oblique or back views where the fixed 135-degree and 225-degree penalty is currently active.
  • The same progressive ordering could be ported to video or 4D editing: choose the frame where an edit is most readable, edit that frame, then propagate to temporally sparse frames, with a refiner to heal occlusions.
  • Because Appendix A.2 concedes that the quantitative metrics are 2D projections and that view consistency is not well captured numerically, the claimed consistency gain deserves a 3D-native check, such as comparing edited Gaussians across many novel viewpoints or measuring surface continuity of the edited region.
  • The paper's Appendix C limitation that training is heavy suggests a targeted ablation: freeze the backbone and train only the small MoVE-LoRA matrices plus the refiner, to see how much of the reported gain comes from the progressive ordering rather than from extra fine-tuning.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes Pro3D-Editor, a text-guided 3D editing framework built on a 'progressive-views paradigm' in which a Primary-view Sampler scores rendered views by editing salience (Eq. 1), Key-view Render propagates the edit from the primary view to other key views via Mixture-of-View-Experts LoRA (Eqs. 2-4), and Full-view Refiner repairs newly rendered views to refine the 3D Gaussian Splatting object (Eq. 5). The authors report quantitative gains over MVEdit, 3DAdapter, Tailor3D, and LGM, including a 47.4% LPIPS improvement and a 9.7% DINO-I improvement, and support the design with component ablations, qualitative comparisons, GPTEval3D preferences, and a small human ranking study.

Significance. The paradigm is conceptually appealing: directing semantic flow from the most editing-salient view to other views is a plausible mechanism for reducing cross-view conflicts, and the module-level ablations (Table 2) and qualitative figures support the internal logic of the design. The paper ships with extensive qualitative materials, including orbiting videos, and the user study, though small, is a useful complement to the image metrics. If the quantitative evidence were strengthened, the method would be a meaningful contribution to the 3D editing literature.

major comments (5)
  1. [4.2 / A.2] Table 1 and Appendix A.2: the two headline metrics, LPIPS and DINO-I, are computed between the edited result and the original object, so they reward identity preservation rather than edit accuracy. A method that applies no edit would achieve perfect scores on both, and the observed CLIP-T gain over the best baseline is only 1.7% (0.299 vs. 0.304). The 47.4% LPIPS and 9.7% DINO-I gains are therefore compatible with the hypothesis that Pro3D-Editor changes the object less than the baselines rather than editing more precisely; the paper needs an edit-region-specific metric (e.g., DINO-I restricted to the edited mask) or a human-judged measure of whether the intended attribute changed.
  2. [A.3 / 4.2] The comparison protocol states that 'the multi-views for Tailor3D and LGM are from our method.' Feeding the baselines the authors' own edited multi-views measures the baselines' ability to reconstruct from those particular views, not their end-to-end editing performance, and it removes the baselines' own multi-view editing from the comparison. All baselines should either use their own edited multi-views or a common neutral input protocol, and the choice must be justified.
  3. [4.1 / Table 1] The evaluation uses 6 objects and 15 prompts, without error bars, significance tests, or a per-prompt breakdown. FID and FVD are distribution-level metrics that require many samples; 72 renders per object is a small sample, and the reported gains could be within run-to-run variance. The authors should report per-random-seed or per-prompt variance or use paired significance tests.
  4. [4.3 / A.2] The paper explicitly states that the quantitative metrics 'fall short of accurately reflecting' view consistency, yet the central claim of the paper is spatial consistency. Consequently, the quantitative tables do not directly support the main claim; the view-consistency evidence consists of qualitative figures and the small user study of Table 3 (8 volunteers, 2 objects each). A dedicated multi-view consistency metric (e.g., cross-view correspondence error) or a substantially larger human study is needed.
  5. [3.1, Eq. (1)] The Primary-view Sampler's score is a hand-crafted combination of softmax-normalized CLIP similarities with a penalty weight alpha=0.5 applied to views at relative azimuths 135 and 225 degrees. The only validation is the two score distributions in Fig. 3 and the single ablation ID-0 vs. ID-1 in Table 2. Since all downstream propagation depends on this choice, the paper should include a sensitivity analysis over alpha and an explanation or test of why 135/225 degrees are the correct penalty locations for the MV-Adapter viewpoint layout.
minor comments (6)
  1. [3.1, Eq. (1)] P(I_c, y_s, y_e)_p is used before it is defined; define the subtraction P(I_c,y_e)_p - P(I_c,y_s)_p before the equation.
  2. [3.2.2, Eq. (4)] M_e is called a binary mask but is used with a linear interpolation; clarify whether M_e is continuous or binarized and how it is computed from 'comparing generated results with the original multi-view images'.
  3. [A.2] 'FID assesses the overall visual similarity between the edited result and the original object' is inaccurate; FID measures distribution distance, not similarity. Also, the description of PSNR as 'reflects changes in detail' is vague.
  4. [4.1 / A.1] Section 4.1 reports 'about 1.5 hours' for the entire editing process, while Appendix A.1 reports 45 minutes for fine-tuning and 45 minutes for 3D editing/refining; reconcile these time budgets.
  5. [2] Preditor3D [5] is cited as a parallel multi-views method but is not included in the quantitative comparison; if it is a direct baseline, the omission should be justified.
  6. [Fig. 1] The subfigures (a)-(e) are referenced in an order that does not match their visual placement, making the paradigm illustration hard to follow.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported gains are measured empirical results, not derived from the paper's own inputs.

full rationale

Pro3D-Editor is an empirical systems paper; its central claims are supported by measured comparisons (Table 1, Table 2, Fig. 5) and ablations, not by a mathematical derivation that can collapse into its own inputs. No fitted parameter is renamed as a prediction: alpha and lambda are fixed hyperparameters (alpha=0.5, lambda=0.5), and the reported LPIPS/DINO-I/CLIP-T numbers are computed on rendered evaluation views after the full pipeline, not optimized to produce those numbers. The Primary-view Sampler's CLIP-based scoring (Eq. 1) does use the same embedding family as the CLIP-T metric, but that is not a reduction: the sampler produces a single pre-edit ranking over input views, while CLIP-T is evaluated on the final 3D edits, and the intervening Key-view Render and Full-view Refiner are substantial trainable transforms. The paper's own statement in Appendix A.2 that the 2D metrics "fall short of accurately reflecting" view consistency concerns metric validity and experimental support, not derivation circularity. Similarly, DINO-I being an identity-similarity measure, so that a no-edit result would score highly, is a correctness and benchmark-design concern rather than a circular derivation. There is no load-bearing self-citation, no imported uniqueness claim, and no equation that equals its own input by construction. The derivation chain is therefore self-contained with respect to circularity, even though the strength of the empirical claims can be questioned on other grounds.

Assumptions & free parameters 6 free parameters · 6 assumptions · 3 invented entities

The paper's central claims rest on a small number of hand-set hyperparameters (alpha, lambda, ranks, training lengths), several domain assumptions about CLIP salience and multi-view diffusion behavior, and three newly introduced modules whose evidence is internal to the paper. No code or data is released, so these components cannot currently be independently verified.

free parameters (6)
  • Primary-view Sampler penalty weight alpha = 0.5
    Hand-set in Sec 4.1; controls the penalty for back-view editing information in Eq. (1). No sensitivity analysis is reported.
  • Two-stage inference fusion weight lambda = 0.5
    Hand-set blend between backbone and MoVE-LoRA multi-view attention in Eq. (4); no ablation on lambda.
  • MoVE-LoRA rank r and lora_alpha = r=32, alpha=16
    Architecture hyperparameters chosen in Sec A.1; control capacity of the low-rank adaptation.
  • Number of view experts M = 6
    Set equal to number of key views in Sec 3.2.1; determines decoupling granularity.
  • ControlNet-Tile LoRA rank and training schedule = rank=64, 1800 steps, lr=1e-3
    Hyperparameters for the repair module in Sec 4.1; chosen without reported sensitivity analysis.
  • 3DGS training steps = 10k + 10k
    Leave-one-out and refinement training lengths in Sec 4.1; no ablation on step count.
assumptions (6)
  • domain assumption CLIP similarity is a valid proxy for editing salience
    Invoked in Eq. (1): the Primary-view Sampler scores views by softmax CLIP similarity to the editing text. No independent validation beyond two examples and one ablation.
  • domain assumption The six fixed perspectives (0, 45, 90, 135, 180, 225 degrees) of the multi-view backbone define the relevant key views, and relative 135/225 views are the safety-critical back views
    Used in Sec 3.1 penalty term; depends on the specific data layout of the MV-Adapter backbone and the claim that missing back views hurt editing quality.
  • ad hoc to paper Blocking reverse gradient flow in MoVE-LoRA (A updated only from primary view) prevents cross-view conflicts
    Sec 3.2.1; presented as a design principle with only ablation support, no theoretical or independent empirical backing.
  • domain assumption The multi-view diffusion backbone (MV-Adapter) can be fine-tuned to propagate edits from primary to other key views using only six rendered views of the object
    Central to Sec 3.2.1; assumes the fine-tuning data from a single object is sufficient to learn generalizable view-to-view edit correspondences.
  • domain assumption Repaired novel views from a fine-tuned 2D diffusion model provide structurally consistent supervision for 3DGS
    Sec 3.3; assumes ControlNet-Tile repairs are multi-view consistent enough to train 3DGS without introducing artifacts.
  • domain assumption 3DGS is a suitable representation for localized editing due to its explicit Gaussian structure
    Sec 2 and Sec 3.3; motivated by the iterative nature of 3DGS but not empirically compared against NeRF or triplane representations in this paper.
invented entities (3)
  • Primary-view Sampler
    purpose: Selects the most editing-salient view as the anchor for the whole editing pipeline
    No external validation; the paper provides only internal score-distribution examples and an ablation vs. random sampling.
  • MoVE-LoRA (Mixture-of-View-Experts Low-Rank Adaption)
    purpose: Decouples feature correspondences from the primary view to each key view during multi-view diffusion fine-tuning
    Only internal ablations (shared LoRA vs. MoVE-LoRA) support it; no public code or independent measurements.
  • Full-view Refiner
    purpose: Repairs fragmented 3DGS regions by fusing edited key-view information with repaired novel views
    Evidence is internal: w/o refiner vs. w/ refiner comparisons; no external dataset or code release.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pro3D-Editor : A Progressive-Views Perspective for Consistent and Precise 3D Editing." pith.science (2026). https://pith.science/paper/MPFQZM2J

@misc{pith2026250600512,
  author       = {Pith},
  title        = {Pith review of: Pro3D-Editor : A Progressive-Views Perspective for Consistent and Precise 3D Editing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MPFQZM2J}},
  note         = {Machine review of arXiv:2506.00512}
}
read the original abstract

Text-guided 3D editing aims to precisely edit semantically relevant local 3D regions, which has significant potential for various practical applications ranging from 3D games to film production. Existing methods typically follow a view-indiscriminate paradigm: editing 2D views indiscriminately and projecting them back into 3D space. However, they overlook the different cross-view interdependencies, resulting in inconsistent multi-view editing. In this study, we argue that ideal consistent 3D editing can be achieved through a \textit{progressive-views paradigm}, which propagates editing semantics from the editing-salient view to other editing-sparse views. Specifically, we propose \textit{Pro3D-Editor}, a novel framework, which mainly includes Primary-view Sampler, Key-view Render, and Full-view Refiner. Primary-view Sampler dynamically samples and edits the most editing-salient view as the primary view. Key-view Render accurately propagates editing semantics from the primary view to other key views through its Mixture-of-View-Experts Low-Rank Adaption (MoVE-LoRA). Full-view Refiner edits and refines the 3D object based on the edited multi-views. Extensive experiments demonstrate that our method outperforms existing methods in editing accuracy and spatial consistency.

Figures

Figures reproduced from arXiv: 2506.00512 by the authors.

Figure 1
Figure 1. We propose a novel editing paradigm (top) for text-guided 3D editing. Compared with [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Method overview. Given a 3D object represented by 3DGS, Pro3D-Editor achieves precise 3D editing, which includes three main steps: (a) Primary-view Sampler selects and edits the most editing-salient view as the primary view. (Sec. 3.1); (b) Key-view Render accurately propagates the editing information from the primary view to local regions of the remaining key views. (Sec. 3.2); (c) Full-view Refiner edits and refin… view at source ↗
Figure 3
Figure 3. Score distribution of Primary-view Sampler. It automatically selects the most editing￾salient view as the primary view based on the given 3DGS object and user-provided editing prompt. When editing a tail, the scores exhibit a reasonable bimodal distribution, peaking at the side views. When editing the dress, the scores show a reasonable unimodal distribution, peaking at the front view. 3.2.1 Mixture-of-View-Experts … view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison with existing methods. LGM and Tailor3D fail to preserve the original features, such as the shoes of a little girl doll. MVEdit inconsistently edits a new face on the back of the head. In comparison, Pro3D-Editor achieves accurate and spatially c…
Figure 5
Figure 5. Figure 5: Quantitative comparison using GPTEval3D [40]. The blue segments indicate the selection rate of Pro3D-Editor, while the orange segments represent that of the baseline. A higher selection rate indicates better editing performance of the corresponding method. 5°. We evalu…
Figure 6
Figure 6. Figure 6: Ablation Studies of MoVE-LoRA. The edited multi-views generated with the MoVE￾LoRA exhibit the most precise 3D local editing and superior spatial consistency in the edited regions. w/o repair w/ repair [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Ablation Studies of Full-view Refiner. Full-view Refiner mitigates structural fragmentation and blurriness, which are caused by directly applying sparse multi-view edits to existing 3D objects. an improvement on the DINO-I metric. As shown in [PITH_FULL_IMAGE:figures/…
Figure 8
Figure 8. Figure 8: Qualitative comparison with existing methods. It can be observed that our method achieves precise and high-quality local 3D editing while addressing the issue of spatial inconsistency. tiveness of MoVE-LoRA in enhancing multi-view editing consistency, as it is difficul…
Figure 9
Figure 9. Figure 9: Ablation studies of Primary-view Sampler. When the randomly selected view is not the most editing-salient view, the editing information from this editing-sparse view may fail to propagate effectively to the editing-salient views, leading to spatially inconsistent acros…
Figure 10
Figure 10. Figure 10: Ablation studies of Primary-view Sampler. Since editing-salient views are difficult to be precisely controlled by editing-sparse views, when the randomly selected view is not the most editing-salient view, the other editing-salient views may produce unreasonable editi…
Figure 11
Figure 11. Figure 11: Ablation studies of MoVE-LoRA. Compared with Shared LoRA, our MoVE-LoRA not only better preserves the features of the original multi-views, but also ensures spatial consistency of the editing regions, achieving precise and consistent multi-view editing. Original w/o R…
Figure 12
Figure 12. Figure 12: Ablation studies of Full-view Refiner. Introducing the Full-view Refiner can improve the quality of the final 3D editing results by eliminating some floating discrete Gaussians, addressing fragmentation issues, and ensuring the structural continuity of the edited 3D o…

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

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

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

    cs.CV 2026-08 conditional novelty 6.0 of 10

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

  2. Hunyuan3D-Buffalo 1.0: A Unified Multimodal Model for Scalable 3D Generation, Understanding, and Editing

    cs.CV 2026-08 conditional novelty 6.0 of 10

    A unified 3D multimodal model combines understanding, text-to-3D generation, instruction-guided editing, and part generation in one architecture, trained on an 87M-sample corpus, with claimed state-of-the-art results.

  3. TRACE: High-Fidelity 3D Scene Editing via Tangible Reconstruction and Geometry-Aligned Contextual Video Masking

    cs.CV 2026-04 conditional novelty 6.0 of 10

    TRACE anchors a video-diffusion editor to 3D meshes to perform consistent part-level edits on 3D Gaussian scenes in about 10 minutes per edit.

Reference graph

Works this paper leans on

44 extracted references · 11 canonical work pages · cited by 3 Pith papers

  1. [1]

    Gaussianeditor: Swift and controllable 3d editing with gaussian splatting

    Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21476–21485, 2024

    work page 2024

  2. [2]

    Shap-editor: Instruction-guided latent 3d editing in seconds

    Minghao Chen, Junyu Xie, Iro Laina, and Andrea Vedaldi. Shap-editor: Instruction-guided latent 3d editing in seconds. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 26456–26466, 2024

    work page 2024

  3. [3]

    Generic 3d diffusion adapter using controlled multi-view editing.arXiv preprint arXiv:2403.12032, 2024

    Hansheng Chen, Ruoxi Shi, Yulin Liu, Bokui Shen, Jiayuan Gu, Gordon Wetzstein, Hao Su, and Leonidas Guibas. Generic 3d diffusion adapter using controlled multi-view editing.arXiv preprint arXiv:2403.12032, 2024

    arXiv 2024

  4. [4]

    3d-adapter: Geometry-consistent multi-view diffusion for high-quality 3d generation.arXiv preprint arXiv:2410.18974, 2024

    Hansheng Chen, Bokui Shen, Yulin Liu, Ruoxi Shi, Linqi Zhou, Connor Z Lin, Jiayuan Gu, Hao Su, Gordon Wetzstein, and Leonidas Guibas. 3d-adapter: Geometry-consistent multi-view diffusion for high-quality 3d generation.arXiv preprint arXiv:2410.18974, 2024

    arXiv

  5. [5]

    Preditor3d: Fast and precise 3d shape editing.arXiv preprint arXiv:2412.06592, 2024

    Ziya Erkoç, Can Gümeli, Chaoyang Wang, Matthias Nießner, Angela Dai, Peter Wonka, Hsin-Ying Lee, and Peiye Zhuang. Preditor3d: Fast and precise 3d shape editing.arXiv preprint arXiv:2412.06592, 2024

    arXiv 2024

  6. [6]

    Gaussctrl: Multi-view consistent text-driven 3d gaussian splatting editing

    Jing Wu, Jia-Wang Bian, Xinghui Li, Guangrun Wang, Ian Reid, Philip Torr, and Victor Adrian Prisacariu. Gaussctrl: Multi-view consistent text-driven 3d gaussian splatting editing. InEuropean Conference on Computer Vision, pages 55–71. Springer, 2024

    work page 2024

  7. [7]

    Instructpix2pix: Learning to follow image editing instructions

    Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18392–18402, 2023

    2023

  8. [8]

    Emu edit: Precise image editing via recognition and generation tasks

    Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8871–8879, 2024

    2024

Show all 44 references
  1. [9]

    Ultraedit: Instruction-based fine-grained image editing at scale

    Haozhe Zhao, Xiaojian Shawn Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. Ultraedit: Instruction-based fine-grained image editing at scale. Advances in Neural Information Processing Systems, 37:3058–3093, 2024

    2024

  2. [10]

    V ox-e: Text-guided voxel editing of 3d objects

    Etai Sella, Gal Fiebelman, Peter Hedman, and Hadar Averbuch-Elor. V ox-e: Text-guided voxel editing of 3d objects. InProceedings of the IEEE/CVF international conference on computer vision, pages 430–440, 2023

    2023

  3. [11]

    Objaverse: A universe of annotated 3d objects

    Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects. arXiv preprint arXiv:2212.08051, 2022

    2022 arXiv

  4. [12]

    Objaverse-xl: A universe of 10m+ 3d objects.arXiv preprint arXiv:2307.05663, 2023

    Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram V oleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl V ondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objave...

    2023 arXiv

  5. [13]

    Zero123++: a single image to consistent multi-view diffusion base model.arXiv preprint arXiv:2310.15110, 2023

    Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model.arXiv preprint arXiv:2310.15110, 2023

    2023 arXiv

  6. [14]

    Wonder3d: Single image to 3d using cross-domain diffusion

    Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. InProceedings of the IEEE/CVF conference on computer vision and pattern...

    2024

  7. [15]

    Era3d: high-resolution multiview diffusion using efficient row-wise attention

    Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wei Xue, Wenhan Luo, et al. Era3d: high-resolution multiview diffusion using efficient row-wise attention. Advances in Neural Information Processing Systems, 37:55975–56000, 2024

    2024

  8. [16]

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

    Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image.arXiv preprint arXiv:2309.03453, 2023. 10

    2023 arXiv

  9. [17]

    Mvdiffusion++: A dense high-resolution multi-view diffusion model for single or sparse-view 3d object reconstruction

    Shitao Tang, Jiacheng Chen, Dilin Wang, Chengzhou Tang, Fuyang Zhang, Yuchen Fan, Vikas Chandra, Yasutaka Furukawa, and Rakesh Ranjan. Mvdiffusion++: A dense high-resolution multi-view diffusion model for single or sparse-view 3d object reconstruction. InEuropean Conference on...

    2024

  10. [18]

    Mv-diffusion: Motion-aware video diffusion model

    Zijun Deng, Xiangteng He, Yuxin Peng, Xiongwei Zhu, and Lele Cheng. Mv-diffusion: Motion-aware video diffusion model. InProceedings of the 31st ACM International Conference on Multimedia, pages 7255–7263, 2023

    2023

  11. [19]

    Imagedream: Image-prompt multi-view diffusion for 3d generation.arXiv preprint arXiv:2312.02201, 2023

    Peng Wang and Yichun Shi. Imagedream: Image-prompt multi-view diffusion for 3d generation.arXiv preprint arXiv:2312.02201, 2023

    2023 arXiv

  12. [20]

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

    Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl V ondrick. Zero-1-to-3: Zero-shot one image to 3d object. InProceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023

    2023

  13. [21]

    Mvdream: Multi-view diffusion for 3d generation.arXiv preprint arXiv:2308.16512, 2023

    Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation.arXiv preprint arXiv:2308.16512, 2023

    2023 arXiv

  14. [22]

    Nerf: Representing scenes as neural radiance fields for view synthesis.Communications of the ACM, 65(1):99–106, 2021

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis.Communications of the ACM, 65(1):99–106, 2021

    2021

  15. [23]

    pixelnerf: Neural radiance fields from one or few images

    Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4578–4587, 2021

    2021

  16. [24]

    Lrm: Large reconstruction model for single image to 3d.arXiv preprint arXiv:2311.04400, 2023

    Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d.arXiv preprint arXiv:2311.04400, 2023

    2023 arXiv

  17. [25]

    Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation

    Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. In European Conference on Computer Vision, pages 1–20. Springer, 2024

    2024

  18. [26]

    Gtr: Improving large 3d reconstruction models through geometry and texture refinement.arXiv preprint arXiv:2406.05649, 2024

    Peiye Zhuang, Songfang Han, Chaoyang Wang, Aliaksandr Siarohin, Jiaxu Zou, Michael Vasilkovsky, Vladislav Shakhrai, Sergey Korolev, Sergey Tulyakov, and Hsin-Ying Lee. Gtr: Improving large 3d reconstruction models through geometry and texture refinement.arXiv preprint arXiv:24...

    2024 arXiv

  19. [27]

    Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model.arXiv preprint arXiv:2311.06214, 2023

    Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model.arXiv preprint arXiv:2311.06214, 2023

    2023 arXiv

  20. [28]

    Upfusion: Novel view diffusion from unposed sparse view observations.arXiv preprint arXiv:2312.06661, 2023

    Bharath Raj Nagoor Kani, Hsin-Ying Lee, Sergey Tulyakov, and Shubham Tulsiani. Upfusion: Novel view diffusion from unposed sparse view observations.arXiv preprint arXiv:2312.06661, 2023

    2023 arXiv

  21. [29]

    Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models.arXiv preprint arXiv:2404.07191, 2024

    Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models.arXiv preprint arXiv:2404.07191, 2024

    2024 arXiv

  22. [30]

    3d gaussian splatting for real-time radiance field rendering.ACM Trans

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4):139–1, 2023

    2023

  23. [31]

    Depth-regularized optimization for 3d gaussian splatting in few-shot images

    Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 811–820, 2024

    2024

  24. [32]

    Qwen2.5-vl technical report.arXiv preprint arXiv:2502.13923, 2025

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Han...

    2025 arXiv

  25. [33]

    Mv-adapter: Multi-view consistent image generation made easy.arXiv preprint arXiv:2412.03632, 2024

    Zehuan Huang, Yuan-Chen Guo, Haoran Wang, Ran Yi, Lizhuang Ma, Yan-Pei Cao, and Lu Sheng. Mv-adapter: Multi-view consistent image generation made easy.arXiv preprint arXiv:2412.03632, 2024

    2024 arXiv

  26. [34]

    Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022. 11

    2022

  27. [35]

    Tip-editor: An accurate 3d editor following both text-prompts and image-prompts.ACM Transactions on Graphics (TOG), 43(4):1–12, 2024

    Jingyu Zhuang, Di Kang, Yan-Pei Cao, Guanbin Li, Liang Lin, and Ying Shan. Tip-editor: An accurate 3d editor following both text-prompts and image-prompts.ACM Transactions on Graphics (TOG), 43(4):1–12, 2024

    2024

  28. [36]

    Dge: Direct gaussian 3d editing by consistent multi-view editing

    Minghao Chen, Iro Laina, and Andrea Vedaldi. Dge: Direct gaussian 3d editing by consistent multi-view editing. InEuropean Conference on Computer Vision, pages 74–92. Springer, 2024

    2024

  29. [37]

    Gaussianobject: Just taking four images to get a high-quality 3d object with gaussian splatting.arXiv e-prints, pages arXiv–2402, 2024

    Chen Yang, Sikuang Li, Jiemin Fang, Ruofan Liang, Lingxi Xie, Xiaopeng Zhang, Wei Shen, and Qi Tian. Gaussianobject: Just taking four images to get a high-quality 3d object with gaussian splatting.arXiv e-prints, pages arXiv–2402, 2024

    2024

  30. [38]

    Tailor3d: Customized 3d assets editing and generation with dual-side images.arXiv preprint arXiv:2407.06191, 2024

    Zhangyang Qi, Yunhan Yang, Mengchen Zhang, Long Xing, Xiaoyang Wu, Tong Wu, Dahua Lin, Xihui Liu, Jiaqi Wang, and Hengshuang Zhao. Tailor3d: Customized 3d assets editing and generation with dual-side images.arXiv preprint arXiv:2407.06191, 2024

    2024 arXiv

  31. [39]

    Lgm: Large multi-view gaussian model for high-resolution 3d content creation

    Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. InEuropean Conference on Computer Vision, pages 1–18. Springer, 2024

    2024

  32. [40]

    Gpt-4v(ision) is a human-aligned evaluator for text-to-3d generation

    Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v(ision) is a human-aligned evaluator for text-to-3d generation. InCVPR, 2024

    2024

  33. [41]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 586–595, 2018

    2018

  34. [42]

    Towards accurate generative models of video: A new metric & challenges.arXiv preprint arXiv:1812.01717, 2018

    Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges.arXiv preprint arXiv:1812.01717, 2018

    2018 arXiv

  35. [43]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pag...

    2021

  36. [44]

    Shared LoRA

    Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection.arXiv preprint arXiv:2203.03605, 2022. 12 A Implementation Details and Comparative Experiments A.1 I...

    2022 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.