Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

InterGSEdit: Interactive 3D Gaussian Splatting Editing with 3D Geometry-Consistent Attention Prior

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read InterGSEdit claims that anchoring 3D Gaussian editing on a user-selected key view and lifting cross-attention maps into a 3D prior yields view-consistent, high-fidelity edits in both non-rigid and rigid tasks.

desk verdict Interactive key-view anchoring is a genuinely useful new angle, but weak empirical support and an untested load-bearing assumption keep this from being a convincing SOTA claim. read the letter →

arxiv 2507.04961 v1 pith:4EULR2D4 submitted 2025-07-07 cs.CV

Minghao Wen , Shengjie Wu , Kangkan Wang , Dong Liang This is my paper · ORCID
classification cs.CV
keywords 3DGaussianSplattingeditingmulti-viewconsistencycross-attentionfusiondiffusionmodelsinteractivenon-rigidCLIP-basedviewselectionattentionunprojection
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper attempts to fix a known failure of text-only 3D editing: when a diffusion model edits each rendered view of a 3D Gaussian scene separately, non-rigid changes such as smiles come out different in different views, producing blur and artifacts. The proposed solution, InterGSEdit, lets a user pick one edited key view that captures the desired look, then uses that choice to select semantically similar reference views, builds a geometry-consistent attention prior by unprojecting the diffusion model's cross-attention maps onto the 3D Gaussians, and re-injects that prior into the diffusion process through a learnable fusion gate. If the method works as claimed, it converts a one-shot text-prompt edit into an interactive, view-consistent edit without sacrificing fine detail.

What carries the argument

The load-bearing object is the 3D Geometry-Consistent Attention Prior $\text{GAP}^{3D}$, built by weighted unprojection of 2D cross-attention maps $M_v(p)$ onto Gaussians via $\text{GAP}^{3D}(i)=\sum_{v} \frac{w_v}{\sum_m w_m}\sum_p M_v(p) O_i(p) T_i(p)$, where $O_i(p)$ and $T_i(p)$ are the opacity and transmittance used in splatting rendering. Around it sit two mechanisms: CLIP-based Semantic Consistency Selection (CSCS), which assigns reference views weights $w_v=\exp(-\gamma \Delta s_v)$ based on how closely their CLIP editing direction matches the key view, and the Attention Fusion Network (AFN), which computes a per-layer gate $G_l=\sigma(W_l[\text{Attn}^{2D}_l;\text{Attn}^{3D}_l]+\gamma(t))$ and fuses $\text{Attn}^{3D}_l$ with $\text{Attn}^{2D}_l$, with $\gamma(t)$ decaying across iterations to shift weight from 3D consistency to 2D detail.

What would settle it

Take the same pipeline but randomly permute the pixel positions of the cross-attention maps before the unprojection in Eq. (8), destroying any spatial correspondence while keeping the same image statistics. If the reported consistency gains over the baselines survive, then GAP3D is not the active mechanism; if tooth artifacts and blur return, the attention-prior geometry is carrying the argument.

Watch

Extended reading notes

Core claim

The central claim is that multi-view consistency in 3D Gaussian editing can be anchored by a single user-selected edited image. InterGSEdit computes a CLIP-based similarity weight for every candidate view relative to the key view, uses those weights to fuse cross-attention maps from the reference views into a 3D prior $\text{GAP}^{3D}$, and then projects that prior back into each view to constrain a diffusion editor. A gated Attention Fusion Network blends the 3D-constrained attention with the editor's native 2D attention, with a linearly decaying bias that favors geometry early in denoising and detail late. On the tested scenes the paper reports state-of-the-art CLIP similarity, CTIDS, and CDC scores, and qualitatively shows that facial edits such as smiles no longer generate tooth artifacts across views.

Load-bearing premise

The method assumes that a diffusion model's cross-attention maps, when mathematically projected onto the 3D Gaussians and back out into other views, encode where each view's edit should go, so that constraining attention is equivalent to constraining geometry, rather than just a heuristic that happens to work on faces.

Editorial extensions

If this is right

  • Users can steer the editing degree by selecting a key view instead of repeatedly re-prompting with text.
  • Non-rigid edits such as facial expressions become view-consistent because inconsistent reference views are down-weighted by CLIP similarity.
  • Rigid edits such as style and appearance changes also improve because the 3D prior anchors geometry while 2D attention preserves detail.
  • The reported CLIP similarity, CTIDS, and CDC improvements over prior baselines support the claim that attention-level 3D constraints help multi-view consistency.

Reading between the lines

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

  • The same weighted unprojection could be extended to video editing by tracking Gaussians across frames, turning key-view selection into temporal consistency control.
  • The gating schedule suggests a general recipe for 3D-aware editors: enforce geometry constraints early in denoising and relax them late for detail, independent of the interactive component.
  • Because CSCS relies on CLIP directional similarity, a testable variant would compare it against human-labeled semantic similarity to see whether selection noise matters.
  • The method leaves open what happens when a user selects a key view whose edit conflicts with the underlying 3D geometry, such as an open-mouth smile on a closed-mouth mesh; mapping that failure boundary would delimit the approach.
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 / 5 minor

Summary. InterGSEdit proposes an interactive 3D Gaussian Splatting editing framework. The user selects one or more key edited views; a CLIP-based Semantic Consistency Selection (CSCS) ranks other edited views by CLIP direction similarity and assigns soft weights; the cross-attention maps of the selected reference views are unprojected onto 3D Gaussians via a weighted splatting operation to form a GAP3D; this prior is re-projected into each target view and fused with the diffusion model's 2D cross-attention by an Attention Fusion Network (AFN) with a time-decaying gate and a KL loss. Experiments on the IN2N datasets compare against IGS2GS, GSEditor, DGE, and a reimplemented VcEdit, with ablations of CSCS and AFN. The paper claims state-of-the-art CLIP similarity, CTIDS, and CDC.

Significance. If the core mechanism holds, InterGSEdit offers a practical way to anchor ambiguous text-guided edits to a user-chosen view, which is a genuine improvement in controllability for non-rigid 3D editing. The GAP3D construction is a natural and interesting extension of VcEdit's attention-unprojection idea, and the AFN gating schedule is a reasonable response to the fidelity-versus-consistency trade-off. The paper also provides qualitative comparisons and ablation studies that illustrate the target artifacts. However, the quantitative evidence is not yet sufficient: the reported numbers are internally inconsistent between Table 1 and Table 2, no variance or per-scene results are given, the VcEdit baseline is an unofficial reimplementation, and the evaluation metrics share the same CLIP embedding space used by the reference-view selection. These issues directly affect the headline SOTA claim.

major comments (5)
  1. [§4.1.2, Eq. (8)] The load-bearing step in the paper is Eq. (8), where cross-attention maps M_v(p) from different reference views are unprojected onto Gaussians and summed into a single prior GAP3D. The paper never specifies which cross-attention layer, which text-token aggregation, or which denoising timestep produces M_v, and it does not show that the unprojected maps from different views agree at corresponding Gaussians. Because the method targets non-rigid regions, where attention can change with expression and tooth visibility, the assumption that these maps are stable measurements of the same 3D content is not self-evident; if they disagree, the sum in Eq. (8) averages away the fine-grained semantics the method aims to preserve. A quantitative single-view stability or cross-view agreement study is needed to validate the GAP3D mechanism.
  2. [§5.3, Table 1 and §5.4, Table 2] Table 1 reports Ours(InfEdit) with CTIDS 0.1531, whereas Table 2, which presumably evaluates the same full configuration ("With AFN"), reports CTIDS 0.2738 on the same metric. The manuscript gives no explanation for this large discrepancy (different scenes, different averaging set, or different runs), and neither table reports standard deviations or per-scene values. Without this information, the claimed improvements in CLIP similarity, CTIDS, and CDC over IGS2GS, GSEditor, DGE, and VcEdit in Table 1 are not statistically grounded.
  3. [§5.4, Table 2] Section 5.4 and Table 2 show that the "Only 2D Attention" ablation reaches CTIDS 0.2835, higher than the full method's 0.2738; the text explains this as expected because full 2D attention edits more extensively, but CTIDS is one of the three headline metrics used to claim SOTA in Section 5.3. The paper should state which metric is primary and provide statistical tests or per-sample values showing that the full method improves CTIDS when consistency is accounted for, and it should discuss the trade-off between edit fidelity and geometric consistency quantitatively rather than only qualitatively.
  4. [§5.2, Table 1] Section 5.2 states that VcEdit's code is not released, so the authors reimplemented its CCM module and evaluated it with InfEdit as the diffusion backbone. A reimplementation may differ from the original method in details that materially affect CLIP similarity, CTIDS, and CDC. The paper should either report the official VcEdit numbers from the original paper, provide a validation that their reimplementation reproduces those numbers, or clearly label the row as "VcEdit (reimplemented)" and discuss potential differences.
  5. [§4.1.1 and §5.1] The CSCS selection rule (Eqs. (3)-(7)) ranks reference views by CLIP direction similarity to the key view, while the quantitative evaluation metrics (CLIP Similarity, CTIDS, CDC) are all built from the same CLIP embedding space. This creates a risk that part of the measured gain comes from selecting views that are CLIP-similar rather than from the 3D attention prior itself. A concrete safeguard would be to add a non-CLIP consistency metric (e.g., pixel-space rendering consistency or a human perceptual study) and to ablate CSCS with a non-CLIP selection rule.
minor comments (5)
  1. [Section 6 and Section 1] There are typographical errors: "consisntency" and "Piror" in the conclusion and "Moverover" in the introduction.
  2. [Eqs. (7) and (9)] The symbol gamma is used both for the temperature coefficient in Eq. (7) and for the time-dependent bias gamma(t) in Eq. (9); the latter also introduces alpha without giving its value or schedule. Please disambiguate the notation and report the hyperparameter values.
  3. [Eq. (8)] The normalization factor in Eq. (8) is written as w_v / sum_m w_m, but the summation index m is not defined in the equation; clarify whether the normalization is per-Gaussian and how views with no contribution to a given Gaussian are handled.
  4. [Eq. (11)] The KL divergence is written as LKL(Attn3D || Attn2D), but the text says the constraint forces Attn2D to converge to Attn3D; specify the exact direction and whether the attention maps are normalized as distributions before computing the KL term.
  5. [Section 4.2] The statement "the weights W_l and a Gating Module are learned" is unclear because the gating module appears to be exactly W_l with a sigmoid; clarify the training procedure for W_l during the diffusion editing process and how it interacts with the 3DGS optimization in Eq. (11).

Circularity Check

1 steps flagged · score 6.0 of 10

CTIDS/CDC gains are partly built into the CLIP-based reference-view selection; other evidence is independent.

  1. fitted input called prediction [Sec. 4.1.1 (Eqs. 5-7) and Sec. 5.1 (Evaluation Metrics); propagated via Eq. (8)-(11)]
    "For a view v, we compute its alignment score as sv = D(∆Iv, ∆T). Using skey as an anchor, we can calculate deviation value ∆sv = |sv − skey| ... wv = exp(−γ∆sv) ... CTIDS [3] is calculated as the cosine similarity between the textual embedding difference ∆T = Etxt_CLIP(Tedit)−Etxt_CLIP(Tsrc) and the corresponding image embedding difference ∆I = Eimg_CLIP(Iedit)−Eimg_CLIP(Isrc)."

    The per-view alignment score s_v in Eq. (5) is the same cosine between CLIP image-edit direction and CLIP text-edit direction that defines CTIDS in Sec. 5.1. CSCS uses s_v (Eqs. 6-7) to select and weight reference views; Eq. (8) averages their attention maps into GAP3D; and AFN (Eqs. 9-10) plus the KL term (Eq. 11) force every edited view's cross-attention toward this weighted average. Therefore the CTIDS and CDC gains are not independent measurements: the output is constructed to follow reference views that were pre-selected for high, mutually consistent values of the very CLIP-direction quantities being scored. The reported improvement on those metrics is partly forced by the selection rather than demonstrated by an external test.

full rationale

The concrete circular step is the identity between the CSCS selection score (Eq. 5) and the CTIDS evaluation metric, combined with the propagation of the selected views' attention through GAP3D (Eq. 8) and AFN (Eqs. 9-11) to all output views. This makes the CTIDS and CDC claims partially self-referential. The paper's CLIP Similarity metric is not identical to the selection criterion, and the qualitative comparisons provide some independent perceptual support, so the circularity is partial rather than total. I did not count the lack of code for VcEdit or the inconsistent CTIDS numbers between Table 1 (0.1531) and Table 2 (0.2738) as circularity; those are correctness and reproducibility concerns. No load-bearing self-citations or imported uniqueness theorems appear in the paper.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The central claim rests on a handful of unstated hyperparameters, a user-interaction assumption, and a geometric interpretation of attention maps. None of these are externally validated, and the evaluation metrics align closely with the optimization targets.

free parameters (5)
  • gamma (temperature coefficient in CSCS weight) = not reported
    Controls the exponential decay in Eq. (7) for reference view weighting; no sensitivity analysis is provided.
  • alpha (decay bias constant in AFN) = not reported
    Scales the linearly decaying bias term gamma(t) in Eq. (9); value is not disclosed.
  • lambda_2D, lambda_3D (loss weights) = not reported
    Weights for editing loss and KL divergence term in Eq. (11); values are not given.
  • K (number of reference views) = not reported
    Top-K selection in Eq. (6); the value of K is not specified.
  • number of sampled views and optimization iterations = 20 random views, 800-1200 iterations
    Experimental settings that affect the statistical behavior of the results, but are not varied or justified.
assumptions (4)
  • domain assumption CLIP embeddings encode editing direction as a linear difference in embedding space
    Used in Eqs. (3)-(5) to define the alignment scores skey and sv. Assumes differences between image/text embeddings correspond to semantic editing changes.
  • domain assumption Cross-attention maps in the diffusion U-Net are spatially aligned with image content and can be unprojected onto 3D Gaussians
    Required for Eq. (8) to produce a meaningful GAP3D. No formal guarantee is given; the paper relies on empirical alignment between attention and geometry.
  • domain assumption The user-selected key view is a reliable anchor for semantic consistency
    The entire CSCS selection and GAP3D construction depends on the key view being representative of the user's intention. This is an interaction design assumption.
  • standard math Standard 3DGS rendering and pre-trained diffusion editing assumptions hold
    Uses the 3DGS rendering equation (Eq. 1) and the InfEdit diffusion backbone as given, without re-deriving them.
invented entities (2)
  • 3D Geometry-Consistent Attention Prior (GAP3D)
    purpose: A per-Gaussian attention score field used to constrain multi-view consistency during diffusion editing
    Constructed from unprojected 2D attention maps of selected reference views; no external validation beyond the paper's own experiments.
  • 3D-constrained attention map (Attn3D)
    purpose: Projection of GAP3D to image space for fusion with 2D cross-attention
    Defined as R(GAP3D, v); same epistemic status as GAP3D, no independent falsifiable handle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of InterGSEdit: Interactive 3D Gaussian Splatting Editing with 3D Geometry-Consistent Attention Prior." pith.science (2026). https://pith.science/paper/4EULR2D4

@misc{pith2026250704961,
  author       = {Pith},
  title        = {Pith review of: InterGSEdit: Interactive 3D Gaussian Splatting Editing with 3D Geometry-Consistent Attention Prior},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4EULR2D4}},
  note         = {Machine review of arXiv:2507.04961}
}
abstract

3D Gaussian Splatting based 3D editing has demonstrated impressive performance in recent years. However, the multi-view editing often exhibits significant local inconsistency, especially in areas of non-rigid deformation, which lead to local artifacts, texture blurring, or semantic variations in edited 3D scenes. We also found that the existing editing methods, which rely entirely on text prompts make the editing process a "one-shot deal", making it difficult for users to control the editing degree flexibly. In response to these challenges, we present InterGSEdit, a novel framework for high-quality 3DGS editing via interactively selecting key views with users' preferences. We propose a CLIP-based Semantic Consistency Selection (CSCS) strategy to adaptively screen a group of semantically consistent reference views for each user-selected key view. Then, the cross-attention maps derived from the reference views are used in a weighted Gaussian Splatting unprojection to construct the 3D Geometry-Consistent Attention Prior ($GAP^{3D}$). We project $GAP^{3D}$ to obtain 3D-constrained attention, which are fused with 2D cross-attention via Attention Fusion Network (AFN). AFN employs an adaptive attention strategy that prioritizes 3D-constrained attention for geometric consistency during early inference, and gradually prioritizes 2D cross-attention maps in diffusion for fine-grained features during the later inference. Extensive experiments demonstrate that InterGSEdit achieves state-of-the-art performance, delivering consistent, high-fidelity 3DGS editing with improved user experience.

Figures

Figures reproduced from arXiv: 2507.04961 by the authors.

Figure 1
Figure 1. Results of a “make him smile” editing task. GSEditor [6] randomly select a subset of edited views as reference view to generate the 3DGS results, leading to blur synthesis due to multi-view inconsistency, such as tooth artifacts in this example. In contrast, our InterGSEdit framework allows the user to select the preferred key views from the edited views to guide the multi-view editing. Here, we illustrate two key v… view at source ↗
Figure 2
Figure 2. Our InterGSEdit framework mainly comprises two components. (a) 3D Geometry-Consistent Attention Prior (GAP3D) Construction. User-specified key views serve to select reference views with high semantic consistency. We employ a CLIP-based Semantic Consistency Selection (CSCS) strategy to select semantically consistent reference views and then utilize their cross-attention maps to construct GAP3D (Sec. 4.1). (b) Adaptiv… view at source ↗
Figure 4
Figure 4. Illustration of GAP3D construction. Based on the attention map Mv corresponding to the view v and the associated camera pose, precise correspondences between image pixels and GS points are established. Each view has a similarity weight wv that is calculated from Eq. (7) to construct the 3D attention prior. Here, only two views are shown for illustration. same textual guidance if ∆sv is close to 0. We select the top-… view at source ↗
Figures from the paper (3 more)
Figure 3
Figure 3. Figure 3: Visualization of CLIP-based Semantic Consistency Selection. The left image is the key view, whose editing direction is quantified by its alignment score skey. Reference views with lower deviation value ∆sv = |sv − skey| indicate that their editing changes are more cons…
Figure 5
Figure 5. Figure 5: Qualitative comparison with IGS2GS [30], GSEditor [6], DGE [4] and VcEdit [33] with CCM module. Our InterGSEdit framework achieves high-quality editing for both non-rigid and rigid tasks, demonstrating strong fidelity to textual instructions, and precise geometry prese…
Figure 6
Figure 6. Figure 6: Ablation studies of our CLIP-based Semantic Consistency Selection (CSCS) and Attention Fusion Network (AFN). In the version “W/O CSCS”, non-rigid editing results, such as facial expression editing, exhibit serious tooth artifacts due to inconsistent features among diff…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. 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

43 extracted references · 34 canonical work pages · cited by 1 Pith paper

  1. [1]

    Score distillation sampling with learned manifold cor- rective

    Thiemo Alldieck, Nikos Kolotouros, and Cristian Sminchis- escu. Score distillation sampling with learned manifold cor- rective. In European Conference on Computer Vision, pages 1–18. Springer, 2024. 3

    work page 2024

  2. [2]

    In- structpix2pix: Learning to follow image editing instructions

    Tim Brooks, Aleksander Holynski, and Alexei A Efros. In- structpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 18392–18402, 2023. 1, 3, 6, 7

    work page 2023

  3. [3]

    Proedit: Simple pro- gression is all you need for high-quality 3d scene editing

    Jun-Kun Chen and Yu-Xiong Wang. Proedit: Simple pro- gression is all you need for high-quality 3d scene editing. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 3, 6, 7

    work page 2024

  4. [4]

    Dge: Di- rect gaussian 3d editing by consistent multi-view editing

    Minghao Chen, Iro Laina, and Andrea Vedaldi. Dge: Di- rect gaussian 3d editing by consistent multi-view editing. In European Conference on Computer Vision, pages 74–92. Springer, 2024. 1, 3, 4, 7, 8

    work page 2024

  5. [5]

    Tango: Text-driven photorealistic and robust 3d styliza- tion via lighting decomposition

    Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d styliza- tion via lighting decomposition. Advances in Neural Infor- mation Processing Systems, 35:30923–30936, 2022. 3

    work page 2022

  6. [6]

    Gaussianeditor: Swift and control- lable 3d editing with gaussian splatting

    Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xi- aofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and control- lable 3d editing with gaussian splatting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21476–21485, 2024. 1, 2, 3, 4, 6, 7, 8

    work page 2024

  7. [7]

    Progressive3d: Progres- sively local editing for text-to-3d content creation with com- plex semantic prompts

    Xinhua Cheng, Tianyu Yang, Jianan Wang, Yu Li, Lei Zhang, Jian Zhang, and Li Yuan. Progressive3d: Progres- sively local editing for text-to-3d content creation with com- plex semantic prompts. In The Twelfth International Confer- ence on Learning Representations, 2024. 3

    work page 2024

  8. [8]

    Lam3d: Large image-point clouds alignment model for 3d reconstruction from single image

    Ruikai Cui, Xibin Song, Weixuan Sun, Senbo Wang, Weizhe Liu, Shenzhou Chen, Taizhang Shang, Y ANG LI, Nick Barnes, Hongdong Li, et al. Lam3d: Large image-point clouds alignment model for 3d reconstruction from single image. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3

    work page 2024

Show all 43 references
  1. [9]

    Blended- nerf: Zero-shot object generation and blending in existing neural radiance fields

    Ori Gordon, Omri Avrahami, and Dani Lischinski. Blended- nerf: Zero-shot object generation and blending in existing neural radiance fields. In 2023 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), page 2933–2943. IEEE, 2023. 3

    2023

  2. [10]

    Instruct-nerf2nerf: Edit- ing 3d scenes with instructions

    Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Edit- ing 3d scenes with instructions. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 19740–19750, 2023. 3, 6, 7

    2023

  3. [11]

    Customize your nerf: Adaptive source driven 3d scene edit- ing via local-global iterative training

    Runze He, Shaofei Huang, Xuecheng Nie, Tianrui Hui, Lu- oqi Liu, Jiao Dai, Jizhong Han, Guanbin Li, and Si Liu. Customize your nerf: Adaptive source driven 3d scene edit- ing via local-global iterative training. In Proceedings of the IEEE/CVF conference on computer vision and ...

    2024

  4. [12]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020. 3

    2020

  5. [13]

    Avatarclip: zero-shot text- driven generation and animation of 3d avatars

    Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: zero-shot text- driven generation and animation of 3d avatars. ACM Trans- actions on Graphics (TOG), 41(4):1–19, 2022. 3

    2022

  6. [14]

    Instruct 3d-to-3d: Text in- struction guided 3d-to-3d conversion

    Hiromichi Kamata, Yuiko Sakuma, Akio Hayakawa, Masato Ishii, and Takuya Narihira. Instruct 3d-to-3d: Text in- struction guided 3d-to-3d conversion. arXiv preprint arXiv:2303.15780, 2023. 3

    2023 arXiv

  7. [15]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 42(4):1–14, 2023. 1, 3, 6

    2023

  8. [16]

    Threestudio: A modular framework for diffusion-guided 3d generation

    Ying-Tian Liu, Yuan-Chen Guo, Vikram V oleti, Ruizhi Shao, Chia-Hao Chen, Guan Luo, Zixin Zou, Chen Wang, Chris- tian Laforte, Yan-Pei Cao, et al. Threestudio: A modular framework for diffusion-guided 3d generation. ICCV , 2023. 6

    2023

  9. [17]

    Latent consistency models: Synthesizing high- resolution images with few-step inference

    Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high- resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1, 6

    2023 arXiv

  10. [18]

    Text2mesh: Text-driven neural stylization for meshes

    Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 13492– 13502, 2022. 3

    2022

  11. [19]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. Communications of the ACM, 65(1):99–106, 2021. 1, 3

    2021

  12. [20]

    Glide: Towards photorealis- tic image generation and editing with text-guided diffusion models

    Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealis- tic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, p...

    2022

  13. [21]

    Learning transferable visual models from natural language supervi- sion

    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 supervi- sion. In International conference on machine learning, ...

    2021

  14. [22]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 1, 3, 6

    2022

  15. [23]

    U- net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- net: Convolutional networks for biomedical image segmen- tation. In Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, pa...

    2015

  16. [24]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 2250...

    2023

  17. [25]

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

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

    2023

  18. [26]

    Dragdiffusion: Harnessing diffusion models for interactive point-based image editing

    Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Han- shu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogniti...

    2024

  19. [27]

    Meshgpt: Generating triangle meshes with decoder-only transformers

    Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Ta- tiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In Proceedings of the IEEE/CVF conference on computer vision and patte...

    2024

  20. [28]

    Dreamgaussian: Generative gaussian splatting for ef- ficient 3d content creation

    Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for ef- ficient 3d content creation. InThe Twelfth International Con- ference on Learning Representations, 2024. 1

    2024

  21. [29]

    Drag-a-video: Non-rigid video editing with point- based interaction

    Yao Teng, Enze Xie, Yue Wu, Haoyu Han, Zhenguo Li, and Xihui Liu. Drag-a-video: Non-rigid video editing with point- based interaction. arXiv preprint arXiv:2312.02936, 2023. 3

    2023 arXiv

  22. [30]

    Instruct-gs2gs: Editing 3d gaussian splats with instructions, 2024

    Cyrus Vachha and Ayaan Haque. Instruct-gs2gs: Editing 3d gaussian splats with instructions, 2024. 7, 8

    2024

  23. [31]

    Clip-nerf: Text-and-image driven ma- nipulation of neural radiance fields

    Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven ma- nipulation of neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3835–3844, 2022. 3

    2022

  24. [32]

    Gaussianeditor: Editing 3d gaussians delicately with text instructions

    Junjie Wang, Jiemin Fang, Xiaopeng Zhang, Lingxi Xie, and Qi Tian. Gaussianeditor: Editing 3d gaussians delicately with text instructions. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 20902–20911, 2024. 1, 3

    2024

  25. [33]

    View-consistent 3d editing with gaus- sian splatting

    Yuxuan Wang, Xuanyu Yi, Zike Wu, Na Zhao, Long Chen, and Hanwang Zhang. View-consistent 3d editing with gaus- sian splatting. In European Conference on Computer Vision, pages 404–420. Springer, 2024. 1, 3, 4, 7, 8

    2024

  26. [34]

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

    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 edit- ing. In European Conference on Computer Vision, pages 55–

  27. [35]

    Unique3d: High-quality and efficient 3d mesh generation from a single image

    Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaisheng Ma. Unique3d: High-quality and efficient 3d mesh generation from a single image. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3

    2024

  28. [36]

    Neural assets: 3d-aware multi-object scene synthesis with image diffusion models

    Ziyi Wu, Yulia Rubanova, Rishabh Kabra, Drew Hud- son, Igor Gilitschenski, Yusuf Aytar, Sjoerd van Steenkiste, Kelsey Allen, and Thomas Kipf. Neural assets: 3d-aware multi-object scene synthesis with image diffusion models. Advances in Neural Information Processing Systems , 3...

    2025

  29. [37]

    Inversion-free image editing with language-guided dif- fusion models

    Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-free image editing with language-guided dif- fusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 9452– 9461, 2024. 3, 6, 7

    2024

  30. [38]

    Nerf-editing: geometry editing of neural radiance fields

    Yu-Jie Yuan, Yang-Tian Sun, Yu-Kun Lai, Yuewen Ma, Rongfei Jia, and Lin Gao. Nerf-editing: geometry editing of neural radiance fields. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 18353–18364, 2022. 3

    2022

  31. [39]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3836–3847, 2023. 3

    2023

  32. [40]

    Towards text-guided 3d scene composition

    Qihang Zhang, Chaoyang Wang, Aliaksandr Siarohin, Peiye Zhuang, Yinghao Xu, Ceyuan Yang, Dahua Lin, Bolei Zhou, Sergey Tulyakov, and Hsin-Ying Lee. Towards text-guided 3d scene composition. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, ...

    2024

  33. [41]

    Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields

    Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Ze- hao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. InPro- ceedings of the IEEE/CVF Conference on Compu...

    2024

  34. [42]

    Dreameditor: Text-driven 3d scene editing with neural fields

    Jingyu Zhuang, Chen Wang, Liang Lin, Lingjie Liu, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. In SIGGRAPH Asia 2023 Conference Papers, pages 1–10, 2023. 3

    2023

  35. [71]

    Springer, 2024. 1, 3, 4

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.