Robust 3D-Masked Part-level Editing in 3D Gaussian Splatting with Regularized Score Distillation Sampling

Hayeon Kim; Ji Ha Jang; Se Young Chun

REVIEW 5 major objections 5 minor 69 references

Robust 3D-Masked Part-level Editing in 3D Gaussian Splatting with Regularized Score Distillation Sampling

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 RoMaP claims that part-level editing of 3D Gaussian scenes becomes precise and can include drastic changes by combining SH-coded soft-label 3D masks with a regularized score distillation loss anchored to SLaMP-edited images.

desk verdict Solid, well-engineered part-level Gaussian editing paper with two genuinely new components; the quantitative evidence is strong but needs error bars, a corrected TIFA number, and a toned-down 'first' claim before I'd accept. read the letter →

arxiv 2507.11061 v3 pith:C7BOX6DO submitted 2025-07-15 cs.CV cs.AI

Hayeon Kim , Ji Ha Jang , Se Young Chun This is my paper

classification cs.CVcs.AI

keywords 3DGaussianSplattingpart-leveleditingscoredistillationsamplingsphericalharmonicslabelpredictiontext-guidedattentionmapsegmentationscheduledlatentmixinglocal

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

RoMaP sets out to make part-level editing of 3D Gaussian Splatting scenes precise enough that a user can change one part, such as a nose, an eye, or a lampshade, while leaving the rest of the scene intact, even for drastic or unusual changes. The paper identifies two obstacles: 2D part segmentations are inconsistent across views, and score distillation sampling (SDS) is too ambiguous to localize and direct an edit. Its mask module, 3D-GALP, stores each Gaussian's part label as spherical-harmonic coefficients, letting a single Gaussian carry different labels in different views, and refines boundaries with an anchor-based neighbor-consistency loss. Its regularized SDS loss removes the original appearance prior, masks the optimization to the target region, and adds an L1 anchor term pulling toward a SLaMP-edited image. If these claims are correct, 3D asset customization gains a controllable part-level editing tool that works on both reconstructed and generated Gaussians.

What carries the argument

The central machinery is the regularized SDS loss paired with the 3D-GALP mask. 3D-GALP adds a learnable SH label field $\mathbf{r}_i$ to each Gaussian, renders it through differentiable rasterization, fits it to SD3 cross-attention maps with an L1 loss, and refines it with an anchor-based neighbor-consistency loss $L_{GALP}$ using Gaussians chosen by label softness $S_i = H_i \cdot v_i$. The editing loss is $L_{R\text{-}SDS} = \lambda_1 \hat{L}_{SDS}(c^{pr}_\phi, p^{edit}) + \lambda_2 \hat{L}_1(c^{pr}_\phi, \mathrm{SLaMP}(c^{pr}_\phi))$, where the $\hat{L}$ notation means masked losses, $c^{pr}_\phi$ is a render of Gaussians whose color priors have been removed, and SLaMP blends target and original latents with a schedule that sharpens at a chosen timestep $t_s$. The mask confines gradients to the target; prior removal lets the edit break away from the original appearance; the SLaMP anchor gives the SDS a concrete direction.

What would settle it

On a well-reconstructed Gaussian scene that also contains a geometrically complex object such as a Klein bottle, run 3D-GALP with the same segmentation prompts and compare its masks against manual part annotations; if the mIoU on the complex object is near chance, or if editing a labeled part changes Gaussians outside the mask beyond a small tolerance, the SD3-attention premise fails. A more direct measurement is per-view IoU between the extracted attention maps and manual part masks: the paper's editing accuracy should degrade monotonically as that IoU drops.

Watch

Extended reading notes

Core claim

The central claim is that local 3D editing of Gaussian Splatting succeeds when both segmentation and optimization respect 3D geometry. The paper's 3D-GALP module adds a learnable label field $\mathbf{r}_i$ to each Gaussian, renders it as $\mathbf{R}_\phi=\mathrm{SH}(\mathbf{r},\phi)$, and fits it to SD3 cross-attention maps with an L1 loss; it then refines the field with an anchor-based neighbor-consistency loss, choosing anchors by label softness $S_i=H_i v_i$ so that boundary Gaussians with genuinely mixed labels are handled explicitly. The editing step replaces plain SDS with $L_{R\text{-}SDS}=\lambda_1\hat{L}_{SDS}(c^{pr}_\phi,p^{edit})+\lambda_2\hat{L}_1(c^{pr}_\phi,\mathrm{SLaMP}(c^{pr}_\phi))$, where the hats denote masking to the target region, $c^{pr}_\phi$ is a render of Gaussians with color priors removed, and SLaMP is a scheduled latent-mixing 2D editor that supplies a concrete modification direction. The paper states that this combination achieves state-of-the-art local editing on reconstructed and generated Gaussian scenes, including drastic edits such as turning a nose into a croissant while preserving the rest of the identity.

Load-bearing premise

The load-bearing premise is that SD3's cross-attention maps, extracted from real rendered views through a controlled forward ODE, are accurate enough to serve as pseudo-ground-truth part labels; the authors themselves state that blurry or poorly reconstructed scenes, and complex geometry like a Klein bottle, make SD3 misinterpret the scene and produce wrong segmentation and undesired edits.

Editorial extensions

If this is right

Users can edit a single part of a reconstructed or generated 3D scene with a text prompt, including rare or statistically unlikely modifications that 2D editors tend to refuse.
Open-vocabulary parts in complex scenes (handles, beaks, collars, name tags) are segmentable and editable because the masks come from diffusion attention maps rather than a fixed part taxonomy.
Part-level editing can steer text-to-3D generation: generate a base object, then edit it part by part, producing assets that are hard to obtain from a single text prompt.
Masking plus prior removal prevents edits from leaking into neighboring regions, so boundary Gaussians with soft labels change only in the intended direction.

Reading between the lines

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

Beyond the paper's experiments, the SH soft-label treatment is a general solution to multi-view label ambiguity and could be reused for 3D segmentation benchmarks or interactive labeling whenever one primitive straddles a part boundary.
A testable extension is a prompt-adaptive schedule for SLaMP's mixing timestep $t_s$; the paper fixes one sweet spot via CLIP direction and SSIM, but per-part or per-prompt schedules may improve the preservation-edit tradeoff.
Because the weak link is SD3 attention, RoMaP's editing fidelity should track the underlying diffusion model's part awareness; swapping in a stronger part-aware diffusion model is a natural stress test.
The Gaussian prior removal trick suggests plain SDS over-anchors to original appearance; a similar neutral-prior step may improve other SDS-based tasks such as stylization, though the paper does not make that claim.

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.

Desk Editor's Note

read the letter

You should read this paper if you care about part-level editing in 3D Gaussian Splatting. It does what it claims, mostly. The two novel components are real: 3D-GALP, which stores a per-Gaussian label as SH coefficients to model view-dependent soft labels and samples anchors by label softness for neighbor consistency, and SLaMP, which does part-confined 2D editing with scheduled latent blending to produce an L1 anchor for a regularized SDS loss. The reported margins over DGE, GaussianEditor, and GaussCtrl are large, especially on the "hard" edits, and the user study with 260 respondents and statistical tests is solid evidence. The mIoU comparison against LangSplat and LeGaussian on 3D-OVS is also to their credit: 0.559 vs. 0.288 average is a real gap. The ablation shows each added component helps, which supports the internal logic.

The soft spots are real but addressable. No quantitative table carries error bars or seed counts, so the headline margins could partly be noise. There is a concrete inconsistency: IN2N's TIFA is 0.634 in Table 1 but 0.565 in Table S.4. That has to be fixed. The hyperparameter t_s is selected using CLIPdir, which is also a headline evaluation metric, so there is some tuning-to-metric pressure; it is not fatal, but it should be disclosed and, ideally, cross-checked with a held-out set. The claim in Section 2.3 about being "the first to achieve part-level 3D editing" is overbroad given earlier local editing works like Local3D and Progressive3D; it should be softened. The deeper concern, which the paper itself concedes in Section S.4, is that the masks inherit errors from SD3 cross-attention pseudo-labels. The average mIoU of 0.559 is good but not great, and failures on complex geometry and blurry scenes mean the editing quality is bounded by the segmentation quality. The neighbor-consistency loss can smooth errors but cannot remove a systematic attention-map failure. That is a genuine limitation, not a hidden flaw, but it should temper how the SOTA claim is phrased.

Who is this for? Anyone working on 3D Gaussian editing or segmentation, especially people building on SDS and text-guided local changes. It deserves a serious referee. My rec: send it to peer review. The central claim holds up conditionally; what it needs is error bars, a corrected TIFA number, a softer 'first' statement, and an explicit note about how t_s is chosen independent of the test prompts.

Referee Report

5 major / 5 minor

Summary. The paper introduces RoMaP, a framework for text-guided part-level editing of 3D Gaussian Splatting scenes. It has two main components. First, 3D-GALP produces a 3D part mask by optimizing per-Gaussian label fields represented as spherical-harmonic coefficients against SD3 cross-attention pseudo-labels, with an anchor-based neighbor-consistency loss that exploits label softness at part boundaries. Second, a regularized SDS loss combines standard SDS with an L1 anchor term derived from SLaMP, a scheduled latent-mixing 2D part-editing method, together with Gaussian prior removal and robust 3D masking. Experiments on IN2N and NeRF-Art scenes compare against DGE, GaussianEditor, GaussCtrl, and several NeRF editing methods, and report quantitative gains in CLIP, CLIPdir, BLIP-VQA, and TIFA, alongside a user study and additional generation-based comparisons.

Significance. If the reported results are reliable, RoMaP is a practical contribution to 3D Gaussian editing: it directly addresses the view-inconsistency of 2D part segmentation and the ambiguity of SDS, and the qualitative results on drastic part edits such as 'croissant nose' and 'left blue, right green eye' are considerably stronger than those of the compared baselines. The idea of storing soft per-Gaussian part labels as spherical-harmonic coefficients is an original and sensible use of Gaussian properties, and the scheduled latent-mixing strategy for generating anchor images is a reasonable mechanism for confining 2D edits. The paper also provides code, a user study with Friedman and Wilcoxon tests, and a supplementary segmentation evaluation against LangSplat and LeGaussian. However, the empirical claims currently rest on single-run tables without variance information, and one headline number is internally inconsistent; the mask pillar also depends on SD3 attention maps whose errors are acknowledged to cause failures. These issues need to be addressed before the state-of-the-art claim can be considered established.

major comments (5)

[Section 4.2, Tables 1 and S.1–S.4] All quantitative tables report single numbers with no standard deviations, confidence intervals, or numbers of seeds. Because the central claim is state-of-the-art performance, the absence of variance information makes it impossible to assess whether the reported margins, such as the 'hard' category CLIPdir of 0.188 versus DGE's 0.110, are stable or arise from a single favorable run. Please provide per-seed results or, at minimum, bootstrap confidence intervals over the 75 prompts.
[Table 1 versus Table S.4] The TIFA score for IN2N is reported as 0.634 in Table 1, but the average of the per-part TIFA values in Table S.4 for IN2N is 0.565. This inconsistency must be resolved; if the aggregate in Table 1 is computed differently, the procedure should be stated, and all aggregate numbers in Tables 1 and S.1–S.4 should be checked for the same issue.
[Sections 3.2, S.3.2.2, S.4, Eq. (2)] The 3D mask is trained entirely from SD3 cross-attention pseudo-labels A(C_φ), and the same mask is used both to confine the SDS/SLaMP losses and to prevent edits outside the target region. Any systematic error in these attention maps, including the failures acknowledged in S.4 for highly complex geometry and blurry reconstructions, will therefore propagate directly into edit leakage or missing edits. The supplementary segmentation evaluation reports an average mIoU of only 0.559 on four 3D-OVS scenes, and no mask accuracy is reported on the IN2N/NeRF-Art scenes used for the main editing tables. The paper should provide mask-quality evidence on the actual evaluation scenes or an explicit sensitivity analysis that injects mask perturbations and measures the effect on edit fidelity.
[Section S.3.3.1 and Section 4.2] The mixing timestep t_s in SLaMP is selected on CelebAMaskHQ by maximizing CLIPdir while keeping SSIM stable, and CLIPdir is also one of the headline evaluation metrics in Tables 1 and S.1–S.4. This couples hyperparameter selection to the evaluation metric. Please report the sensitivity of the final 3D editing scores to t_s, or tune on a separate validation set and state that the reported metrics correspond to held-out prompts.
[Algorithm 2] The SLaMP pseudocode is not reproducible as written: the second loop reuses the loop variable t from the first loop, the variable i used in the F_t assignment is never defined, and z_noisy is updated both inside the first loop and via scheduler.step in the second loop. Please rewrite the algorithm with distinct iterator names and a single, clearly specified update rule for z_noisy.

minor comments (5)

[Section 3.3, Eq. (3)] The masked losses Rhat L_SDS and Rhat L_1 are not explicitly defined; please state which masks (M_2D, M_3D) enter each term and how the masking is applied to the SDS gradient.
[Section 3.2, Eq. (2)] The values of the anchor count K and the neighbor count k are not reported anywhere; please include them in the implementation details.
[Section S.3.1] The softmax temperature for attention-map extraction is described as adjustable, but the default value used in the editing experiments is not stated; the segmentation experiment in Table S.7 used 0.2, so please clarify whether the same value was used throughout.
[Figure 3] The 'Scheduled weight mask' box in the pipeline figure is not connected to the equation for F_t in Eq. (4); consider annotating the figure to make the correspondence explicit.
[Supplementary, figure captions and text] There are several typos, including 'nerf baseines' in Fig. S.15 and 'quantiative' in Section S.1.2; please proofread the supplementary text.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the core mask and editing derivations are self-contained, with a minor self-referential hyperparameter choice that does not reduce the central claims to their inputs.

full rationale

RoMaP's derivation chain is not circular in the formal sense. The 3D-GALP mask is trained against SD3 cross-attention maps treated as pseudo 2D labels, but its reported quality is validated against human-annotated part segmentations (Table S.7, mIoU 0.559 vs. LangSplat 0.076 and LeGaussian 0.288), so the mask claim has independent external support. The regularized SDS loss (Eq. 3) combines SD3-based SDS with an L1 anchor from SLaMP, all drawn from the same SD3 model family; this creates model-consistency pressure but does not make any predicted quantity equal to an input by construction. The final editing quality is assessed with user studies, CLIP-based metrics, BLIP-VQA, and TIFA, which are external to the fitted losses. One mild self-referential point is that the SLaMP timestep t_s is selected on CelebAMaskHQ using CLIPdir and SSIM (Sec. S.3.3.1), and CLIPdir is also a headline evaluation metric; this is hyperparameter tuning on a held-out image set rather than a fitted parameter renamed as a prediction, so it does not invalidate the reported comparisons. The paper also explicitly states its limitation in Sec. S.4: results depend on SD3's ability to interpret complex or blurry scenes, which is an acknowledged assumption about generalization rather than a circular derivation. No load-bearing self-citations are present; the authors do not rely on their own prior work to justify the core method. Overall, the paper is self-contained against external benchmarks and its central claim does not reduce to its inputs.

Assumptions & free parameters 6 free parameters · 5 assumptions · 1 invented entities

The central claim rests on several empirical assumptions about the quality of SD3 attention maps as pseudo-labels, the differentiability of the SH label field, the effectiveness of the regularized SDS objective, and the validity of the evaluation metrics. No formal derivations are offered; the method is validated only by its reported tables and user study.

free parameters (6)

lambda_1, lambda_2
Weights for SDS and L1 anchor terms in Eq. (3); the authors do not report their values, yet they control the trade-off between global diffusion guidance and localized SLaMP anchoring.
t_s (SLaMP mix timestep) = chosen at SSIM-stable/CLIPdir-high point (Fig. 6)
The sharp blending switch timestep in Eq. (4) is selected by evaluating SSIM and CLIPdir on 100 CelebAMaskHQ images with 25 prompts; this is a fitted hyperparameter.
K (number of anchors) and k (neighbor count)
3D-GALP anchor sampling and neighbor consistency (Eq. 2, Algo. 1) require K and k; not reported.
alpha_base, alpha_last, gamma, eta
Blending and interpolation coefficients in Algo. 2 controlling latent mixing strength; no values given.
softmax temperature for attention map extraction = 0.2 (supplement S.3.2.2)
Post-processing temperature for normalizing SD3 attention maps into pseudo-labels; fixed by hand to 0.2.
SH order for Gaussian labels = 3
Chosen from an ablation (SH=1..4) as the best trade-off; a manual model selection, not derived.

assumptions (5)

domain assumption SD3 cross-attention maps from controlled forward ODE inversion are accurate part-level pseudo-labels.
Used as ground truth in Section 3.2 for the pixel-wise L1 loss on the rendered label field; the paper's limitation (Section S.4) concedes failures on complex geometry and blurry reconstructions.
domain assumption The SH-encoded label field r_i can be optimized via differentiable rasterization with L1 loss and neighbor consistency to converge to a view-consistent 3D segmentation.
Central to 3D-GALP; relies on 3DGS differentiability and the heuristic that neighbor smoothing corrects noisy 2D labels (Eq. 1-2).
domain assumption SDS gradients plus the SLaMP-anchored L1 term and prior removal produce edits confined to the mask.
The regularized SDS loss (Eq. 3) is asserted to give controllable drastic edits; no derivation is provided and ablation shows each component helps empirically.
domain assumption The scheduled latent blending of Eq. (4) preserves context outside the mask while enabling drastic in-mask changes.
SLaMP's sharp transition at t_s is an empirical design; its success is validated only through CLIPdir/B-VQA and the final 3D results.
domain assumption CLIP, CLIPdir, BLIP-VQA, TIFA, and the user study are valid measures of editing alignment and quality.
All headline conclusions in Tables 1-2 rest on these metrics; their alignment with human judgment is presumed (user study is included but percentages are hard to interpret).

invented entities (1)

Per-Gaussian label field r_i, stored as SH coefficients independent evidence
purpose: Enables view-dependent soft part labels for 3D segmentation and is rendered as a segmentation map R_phi.
The field's correctness is tested against manually annotated part segmentation (mIoU on 3D-OVS, Table S.7) and through downstream editing results, providing an external falsifiable handle; but it is a new parameter added to the Gaussian representation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Robust 3D-Masked Part-level Editing in 3D Gaussian Splatting with Regularized Score Distillation Sampling." pith.science (2026). https://pith.science/paper/C7BOX6DO

@misc{pith2026250711061,
  author       = {Pith},
  title        = {Pith review of: Robust 3D-Masked Part-level Editing in 3D Gaussian Splatting with Regularized Score Distillation Sampling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C7BOX6DO}},
  note         = {Machine review of arXiv:2507.11061}
}

read the original abstract

Recent advances in 3D neural representations and instance-level editing models have enabled the efficient creation of high-quality 3D content. However, achieving precise local 3D edits remains challenging, especially for Gaussian Splatting, due to inconsistent multi-view 2D part segmentations and inherently ambiguous nature of Score Distillation Sampling (SDS) loss. To address these limitations, we propose RoMaP, a novel local 3D Gaussian editing framework that enables precise and drastic part-level modifications. First, we introduce a robust 3D mask generation module with our 3D-Geometry Aware Label Prediction (3D-GALP), which uses spherical harmonics (SH) coefficients to model view-dependent label variations and soft-label property, yielding accurate and consistent part segmentations across viewpoints. Second, we propose a regularized SDS loss that combines the standard SDS loss with additional regularizers. In particular, an L1 anchor loss is introduced via our Scheduled Latent Mixing and Part (SLaMP) editing method, which generates high-quality part-edited 2D images and confines modifications only to the target region while preserving contextual coherence. Additional regularizers, such as Gaussian prior removal, further improve flexibility by allowing changes beyond the existing context, and robust 3D masking prevents unintended edits. Experimental results demonstrate that our RoMaP achieves state-of-the-art local 3D editing on both reconstructed and generated Gaussian scenes and objects qualitatively and quantitatively, making it possible for more robust and flexible part-level 3D Gaussian editing. Code is available at https://janeyeon.github.io/romap.

Figures

Figures reproduced from arXiv: 2507.11061 by the authors.

**Figure 1.** Enhanced controllability in 3D Gaussian part-level editing achieved with RoMaP, surpassing prior arts. RoMaP enables highly controllable and localized part-level edits, allowing even for unconventional modifications such as ‘emerald nose’ or modifications requiring a high-level controllability such as ‘blue left eye, right green eye’ while maintaining global consistency. In contrast, existing baselines perform well … view at source ↗

**Figure 2.** Limitations of prior local 3D editing methods leveraging 2D part level segmentation and edits. Although existing 3D editing methods excel in instance level editing, they struggle with part level editing as part segmentation [28] (for ‘eye’) lacks view consistency, and 2D editing [2] often misplaces changes, turning a wall green instead of the left eye. In contrast, our method achieves accurate 3D eye segmentation wi… view at source ↗

**Figure 3.** Overall pipeline of RoMaP. RoMaP first segments 3D Gaussian using 3D-GALP, leveraging the soft-label properties of Gaussians to address the intricacies of part-level segmentation. With anchors consisting of both label-consistent and inconsistent Gaussians, we refine 3D segmentation considering locality with neighboring Gaussians. Then, in local 3D editing, we first remove Gaussian priors and introduce a new modific… view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Effectiveness of label softness-based anchor sampling. By applying 3D loss with anchors sampled based on label softness, we observe that differentiation of boundaries between parts is much more precise compared to random sampling. with high soft-label properties and …

**Figure 5.** Figure 5 [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Experiments on the effect of ts. ts controls the extent of deviation from the original. We set ts to induce drastic changes in the target region while preserving the surrounding identity. masked loss leveraging M3D and M2D to restrict changes only to intended regions. …

**Figure 7.** Figure 7: Enhanced controllability in 3D asset generation with RoMaP. Our approach enables precise manipulation of specific 3D parts. As shown above, RoMaP provides diverse control over multiple narrow regions within a single 3D object, allowing deformations in targeted areas li…

**Figure 8.** Figure 8: Comparison results The results of comparing our methodology with various reconstruction-based 3D editing methods and text-to-3D generation approaches are presented. In the reconstructed scene, our method enables drastic changes in very narrow regions, breaking the exis…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

69 extracted references · 54 canonical work pages

[1]

An examination of the robustness of reference-free image captioning evaluation metrics.ACL Anthology, 2023

Saba Ahmadi and Aishwarya Agrawal. An examination of the robustness of reference-free image captioning evaluation metrics.ACL Anthology, 2023. 10

work page 2023
[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. InCVPR, 2023. 2, 3, 7, 10, 15, 16

work page 2023
[3]

Effi- cient geometry-aware 3d generative adversarial networks

Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Effi- cient geometry-aware 3d generative adversarial networks. In CVPR, 2022. 4

work page 2022
[4]

Mvdrag3d: Drag-based creative 3d editing via multi-view generation-reconstruction priors.arXiv preprint arXiv:2410.16272, 2024

Honghua Chen, Yushi Lan, Yongwei Chen, Yifan Zhou, and Xingang Pan. Mvdrag3d: Drag-based creative 3d editing via multi-view generation-reconstruction priors.arXiv preprint arXiv:2410.16272, 2024. 3, 5, 6

arXiv 2024
[5]

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

Minghao Chen, Iro Laina, and Andrea Vedaldi. Dge: Direct gaussian 3d editing by consistent multi-view editing.ECCV,

work page
[6]

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. InCVPR, 2024. 2, 3, 5, 6, 7, 10, 11, 13, 26

work page 2024
[7]

Text-to-3d using gaussian splatting

Zilong Chen, Feng Wang, Yikai Wang, and Huaping Liu. Text-to-3d using gaussian splatting. InCVPR, 2024. 7, 13

work page 2024
[8]

Cascadepsp: Toward class-agnostic and very high- resolution segmentation via global and local refinement

Ho Kei Cheng, Jihoon Chung, Yu-Wing Tai, and Chi-Keung Tang. Cascadepsp: Toward class-agnostic and very high- resolution segmentation via global and local refinement. In CVPR, 2020. 13

work page 2020

Show all 69 references

[9]

Progressive3d: Progres- sively local editing for text-to-3d content creation with com- plex semantic prompts.ICLR, 2023

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.ICLR, 2023. 2, 3, 10

2023
[10]

Cogview: Mastering text-to-image gen- eration via transformers.NeurIPS, 2021

Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image gen- eration via transformers.NeurIPS, 2021. 2

2021
[11]

Density estimation using real nvp.ICLR, 2016

Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp.ICLR, 2016. 2

2016
[12]

Vica-nerf: View- consistency-aware 3d editing of neural radiance fields

Jiahua Dong and Yu-Xiong Wang. Vica-nerf: View- consistency-aware 3d editing of neural radiance fields. NeurIPS, 2023. 2, 3, 6, 7, 10, 11

2023
[13]

Scaling recti- fied flow transformers for high-resolution image synthesis

Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling recti- fied flow transformers for high-resolution image synthesis. InICML, 2024. 2, 10, 15

2024
[14]

Stylegan-nada: Clip- guided domain adaptation of image generators

Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip- guided domain adaptation of image generators. 2022. 6, 10, 15

2022
[15]

Stable- dreamer: Taming noisy score distillation sampling for text- to-3d.arXiv preprint arXiv:2312.02189, 2023

Pengsheng Guo, Hans Hao, Adam Caccavale, Zhongzheng Ren, Edward Zhang, Qi Shan, Aditya Sankar, Alexan- der G Schwing, Alex Colburn, and Fangchang Ma. Stable- dreamer: Taming noisy score distillation sampling for text- to-3d.arXiv preprint arXiv:2312.02189, 2023. 3, 5, 6

2023 arXiv
[16]

threestudio: A unified framework for 3d content generation

Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram V oleti, Guan Luo, Chia-Hao Chen, Zi- Xin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation. https://github.com/threestudio- project/ threestudio, 2023. 10

2023
[17]

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. InCVPR, 2023. 2, 3, 6, 7, 10, 11, 13, 25

2023
[18]

Prompt-to-prompt image editing with cross attention control.ICLR, 2022

Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control.ICLR, 2022. 2

2022
[19]

Denoising diffu- sion probabilistic models.NeurIPS, 2020

Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models.NeurIPS, 2020. 2

2020
[20]

Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering

Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Os- tendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. InCVPR, 2023. 6, 10

2023
[21]

T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation

Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xi- hui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. 2023. 6, 10, 15

2023
[22]

Local 3d editing via 3d distillation of clip knowledge

Junha Hyung, Sungwon Hwang, Daejin Kim, Hyunji Lee, and Jaegul Choo. Local 3d editing via 3d distillation of clip knowledge. InCVPR, 2023. 4

2023
[23]

Noise-free score distillation.ICLR, 2023

Oren Katzir, Or Patashnik, Daniel Cohen-Or, and Dani Lischinski. Noise-free score distillation.ICLR, 2023. 3, 5

2023
[24]

Imagic: Text-based real image editing with diffusion models

Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In CVPR, 2023. 2

2023
[25]

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

Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 2023. 2, 4

2023
[26]

Lerf: Language embedded radiance fields

Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 19729–19739,
[27]

Unifiedqa-v2: Stronger generalization via broader cross- format training

Daniel Khashabi, Yeganeh Kordi, and Hannaneh Hajishirzi. Unifiedqa-v2: Stronger generalization via broader cross- format training. 2022. 10

2022
[28]

Segment any- thing

Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. InICCV, 2023. 2, 3

2023
[29]

Posterior distil- lation sampling

Juil Koo, Chanho Park, and Minhyuk Sung. Posterior distil- lation sampling. InCVPR, 2024. 3, 6, 7, 10, 11

2024
[30]

Maskgan: Towards diverse and interactive facial image ma- nipulation

Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image ma- nipulation. InCVPR, 2020. 15

2020
[31]

Dreamcou- ple: Exploring high quality text-to-3d generation via rectified flow.arXiv preprint arXiv:2408.05008, 2024

Hangyu Li, Xiangxiang Chu, and Dingyuan Shi. Dreamcou- ple: Exploring high quality text-to-3d generation via rectified flow.arXiv preprint arXiv:2408.05008, 2024. 3

2024 arXiv
[32]

Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation

Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. InICML,
[33]

Dreambeast: Distilling 3d fantastical animals with part-aware knowledge transfer.3DV, 2025

Runjia Li, Junlin Han, Luke Melas-Kyriazi, Chunyi Sun, Zhaochong An, Zhongrui Gui, Shuyang Sun, Philip Torr, and Tomas Jakab. Dreambeast: Distilling 3d fantastical animals with part-aware knowledge transfer.3DV, 2025. 3

2025
[34]

Focaldreamer: Text- driven 3d editing via focal-fusion assembly

Yuhan Li, Yishun Dou, Yue Shi, Yu Lei, Xuanhong Chen, Yi Zhang, Peng Zhou, and Bingbing Ni. Focaldreamer: Text- driven 3d editing via focal-fusion assembly. InAAAI, 2024. 3, 10

2024
[35]

Weakly supervised 3d open- vocabulary segmentation.NeurIPS, 2023

Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open- vocabulary segmentation.NeurIPS, 2023. 13, 14

2023
[36]

Flow straight and fast: Learning to generate and transfer data with rectified flow.ICLR, 2023

Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.ICLR, 2023. 2

2023
[37]

Repaint: Inpainting using denoising diffusion probabilistic models

Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461–11471, 2022. 15

2022
[38]

Latent consistency models: Synthesizing high- resolution images with few-step inference.arXiv preprint arXiv:2310.04378, 2023

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. 2

2023 arXiv
[39]

Nerf: Representing scenes as neural radiance fields for view syn- thesis.ECCV, 2020

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.ECCV, 2020. 2

2020
[40]

Gsedit: Efficient text-guided edit- ing of 3d objects via gaussian splatting.arXiv preprint arXiv:2403.05154, 2024

Francesco Palandra, Andrea Sanchietti, Daniele Baieri, and Emanuele Rodol `a. Gsedit: Efficient text-guided edit- ing of 3d objects via gaussian splatting.arXiv preprint arXiv:2403.05154, 2024. 2, 3

2024 arXiv
[41]

Pytorch: An imperative style, high-performance deep learning library

Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zem- ing Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 10

2019
[42]

Scalable diffusion models with transformers

William Peebles and Saining Xie. Scalable diffusion models with transformers. InCVPR, 2023. 2

2023
[43]

Dreamfusion: Text-to-3d using 2d diffusion

Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Milden- hall. Dreamfusion: Text-to-3d using 2d diffusion. 2023. 10

2023
[44]

Dead- iff: An efficient stylization diffusion model with disentan- gled representations

Tianhao Qi, Shancheng Fang, Yanze Wu, Hongtao Xie, Ji- awei Liu, Lang Chen, Qian He, and Yongdong Zhang. Dead- iff: An efficient stylization diffusion model with disentan- gled representations. InCVPR, 2024. 2

2024
[45]

Langsplat: 3d language gaussian splatting

Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. InCVPR, 2024. 14

2024
[46]

Learn- ing transferable visual models from natural language super- vision

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. InICML, 2021. 4, 6, 10

2021
[47]

Hierarchical text-conditional image gen- eration with clip latents.arXiv preprint arXiv:2204.06125,

Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gen- eration with clip latents.arXiv preprint arXiv:2204.06125,

arXiv
[48]

High-resolution image syn- thesis with latent diffusion models

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. InCVPR, 2022. 2

2022
[49]

Semantic im- age inversion and editing using rectified stochastic differen- tial equations.arXiv preprint arXiv:2410.10792, 2024

Litu Rout, Yujia Chen, Nataniel Ruiz, Constantine Carama- nis, Sanjay Shakkottai, and Wen-Sheng Chu. Semantic im- age inversion and editing using rectified stochastic differen- tial equations.arXiv preprint arXiv:2410.10792, 2024. 3, 13, 15

arXiv
[50]

Deep marching tetrahedra: a hybrid representa- tion for high-resolution 3d shape synthesis.NeurIPS, 2021

Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representa- tion for high-resolution 3d shape synthesis.NeurIPS, 2021. 2

2021
[51]

Language embedded 3d gaussians for open- vocabulary scene understanding

Jin-Chuan Shi, Miao Wang, Hao-Bin Duan, and Shao- Hua Guan. Language embedded 3d gaussians for open- vocabulary scene understanding. InCVPR, 2024. 14

2024
[52]

Emscore: Evaluating video captioning via coarse-grained and fine-grained embed- ding matching

Yaya Shi, Xu Yang, Haiyang Xu, Chunfeng Yuan, Bing Li, Weiming Hu, and Zheng-Jun Zha. Emscore: Evaluating video captioning via coarse-grained and fine-grained embed- ding matching. InCVPR, 2022. 10

2022
[53]

Deep unsupervised learning using nonequilibrium thermodynamics

Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. InICML, 2015. 2

2015
[54]

Llama 2: Open foundation and fine-tuned chat models

Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. 2023. 10

2023
[55]

Textmesh: Gen- eration of realistic 3d meshes from text prompts

Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Gen- eration of realistic 3d meshes from text prompts. In3DVS,
[56]

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

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

2024
[57]

Diffusers: State-of-the-art diffusion models, 2022

Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, William Berman, Yiyi Xu, Steven Liu, and Thomas Wolf. Diffusers: State-of-the-art diffusion models, 2022. 15

2022
[58]

Nerf-art: Text-driven neural radiance fields stylization.TVCG, 2023

Can Wang, Ruixiang Jiang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization.TVCG, 2023. 2, 10, 16

2023
[59]

Sclip: Rethinking self-attention for dense vision-language inference

Feng Wang, Jieru Mei, and Alan Yuille. Sclip: Rethinking self-attention for dense vision-language inference. InECCV,
[60]

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. InCVPR, 2024. 2, 3

2024
[61]

Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction

Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2

2021 arXiv
[62]

Gaussctrl: multi-view consistent text-driven 3d gaussian splatting edit- ing.ECCV, 2024

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.ECCV, 2024. 2, 3, 6, 7, 10, 11, 13

2024
[63]

Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation

Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, 2023. 2

2023
[64]

Gg-editor: Locally editing 3d avatars with multimodal large language model guidance

Yunqiu Xu, Linchao Zhu, and Yi Yang. Gg-editor: Locally editing 3d avatars with multimodal large language model guidance. InACM International Conference on Multimedia,
[65]

Text-to-image rectified flow as plug-and-play priors.arXiv preprint arXiv:2406.03293, 2024

Xiaofeng Yang, Cheng Chen, Xulei Yang, Fayao Liu, and Guosheng Lin. Text-to-image rectified flow as plug-and-play priors.arXiv preprint arXiv:2406.03293, 2024. 3, 7, 13, 15

2024 arXiv
[66]

V ol- ume rendering of neural implicit surfaces.NeruIPS, 2021

Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. V ol- ume rendering of neural implicit surfaces.NeruIPS, 2021. 2

2021
[67]

Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models

Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. InCVPR,
[68]

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 CVPR, 2023. 7

2023
[2024]

2, 3, 6, 7, 10, 11, 13

Pith tools

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

[1] [1]

An examination of the robustness of reference-free image captioning evaluation metrics.ACL Anthology, 2023

Saba Ahmadi and Aishwarya Agrawal. An examination of the robustness of reference-free image captioning evaluation metrics.ACL Anthology, 2023. 10

work page 2023

[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. InCVPR, 2023. 2, 3, 7, 10, 15, 16

work page 2023

[3] [3]

Effi- cient geometry-aware 3d generative adversarial networks

Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Effi- cient geometry-aware 3d generative adversarial networks. In CVPR, 2022. 4

work page 2022

[4] [4]

Mvdrag3d: Drag-based creative 3d editing via multi-view generation-reconstruction priors.arXiv preprint arXiv:2410.16272, 2024

Honghua Chen, Yushi Lan, Yongwei Chen, Yifan Zhou, and Xingang Pan. Mvdrag3d: Drag-based creative 3d editing via multi-view generation-reconstruction priors.arXiv preprint arXiv:2410.16272, 2024. 3, 5, 6

arXiv 2024

[5] [5]

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

Minghao Chen, Iro Laina, and Andrea Vedaldi. Dge: Direct gaussian 3d editing by consistent multi-view editing.ECCV,

work page

[6] [6]

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. InCVPR, 2024. 2, 3, 5, 6, 7, 10, 11, 13, 26

work page 2024

[7] [7]

Text-to-3d using gaussian splatting

Zilong Chen, Feng Wang, Yikai Wang, and Huaping Liu. Text-to-3d using gaussian splatting. InCVPR, 2024. 7, 13

work page 2024

[8] [8]

Cascadepsp: Toward class-agnostic and very high- resolution segmentation via global and local refinement

Ho Kei Cheng, Jihoon Chung, Yu-Wing Tai, and Chi-Keung Tang. Cascadepsp: Toward class-agnostic and very high- resolution segmentation via global and local refinement. In CVPR, 2020. 13

work page 2020