acceptodds
Under review as a conference paper at ICLR 2027

T3HG-Editor: Text-driven 3D Human Garment Editing with Body Priors Embedded in SMPL-X

Abstract

Despite recent advances in 3D Gaussian Editing (3DGE), text-driven 3D human garment editing remains underexplored. Existing methods edit multi-view renderings with 2D techniques and update the underlying Gaussians, often causing geometric distortions and cross-view inconsistencies. SMPL eXpressive (SMPL-X) provides rich human geometry and joint priors to address these challenges. Motivated by this insight, we propose a text-driven 3D human garment editor termed T3HG-Editor, which delivers high-fidelity and garment consistent results by leveraging geometry and joint priors embedded in SMPL-X. Specifically, T3HG-Editor contains three stages, namely obtainment of editable Gaussians, garment consistent editing, and Gaussian updating with overflow pruning. The obtainment of editable Gaussians begins with seeding Gaussians along SMPL-X normals to generate sufficient near surface Gaussians, followed by a 2D mask constraint that precisely localizes the target Gaussians to be edited. The garment consistent editing aggregates tokens corresponding to the same SMPL-X vertex across multiple views and propagates them to their original views, enforcing garment consistency without requiring additional training. Gaussian updating with overflow pruning employs a Signed Distance Function (SDF) defined on SMPL-X to construct a human distance field, which is then integrated with a 2D semantic mask to prune overflowing Gaussians, thus preventing contamination of non-target regions. Experiments on multiple subjects and diverse garment types demonstrate that T3HG-Editor outperforms state-of-the-art methods in both editing quality and garment consistency.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Gaussian splatting, human garment editing, SMPL-X

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_CHjbOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents T3HG-Editor, a text-driven 3D human garment editing framework based on SMPL-X priors. It introduces SMPL-X-guided Gaussian initialization, vertex-based cross-view feature consistency, and geometry-aware Gaussian pruning to enable more consistent garment editing in 3D Gaussian representations. The method is evaluated on reconstructed human Gaussian scenes with text-driven garment editing tasks and demonstrates improved editing quality compared with existing approaches, with notable improvements in non-target region preservation and perceptual garment consistency.

Critical strengths
  1. Leveraging SMPL-X vertex correspondences for cross-view attention is an interesting idea for improving garment consistency.
  2. The method demonstrates promising editing results, particularly in preserving non-target regions, with supporting component-wise ablations.
Critical weaknesses

Although the paper demonstrates strong preservation of non-target regions, the improvement on the edited garment region itself is relatively limited. The reported metrics show only marginal gains on garment regions, and the current evaluation may not fully demonstrate improved garment fidelity and 3D geometric consistency.

Questions for authors

Could the authors provide more direct evaluations of garment fidelity and 3D consistency, such as garment-specific metrics or cross-view geometric consistency measurements?

3 comments
AuthorsOpenReview ↗

Introduction of the Editing Success Rate (ESR) metric and clarification of garment editing effectiveness and cross-view consistency

We sincerely thank the reviewer for this valuable comment. We agree that additional evaluation metrics should be introduced, and we would like to clarify why the previously reported metrics may understate our improvement.

Since garment editing should first and foremost be judged by whether the result follows the instruction, we introduce the Editing Success Rate (ESR) in the rebuttal version, which jointly considers instruction compliance, garment structural completeness and geometric plausibility, and visual realism. Our method achieves the highest ESR and outperforms all compared methods in target-region SSIM (0.5796) and LPIPS (0.1738), even after substantially modifying garment textures and geometric structures, providing stronger evidence of improved garment fidelity and cross-view consistency. Further discussion is provided in Section 4.3 (p. 9) and the Appendix (pp. 18–19).

Moreover, PSNR, SSIM, and LPIPS cannot faithfully reflect the impact of editing failures. When editing fails, the compared methods tend to retain the original garments' appearance and pixel-level textures. As a result, they can obtain favorable image-based similarity scores although the edited results do not conform to the instructions.

AuthorsOpenReview ↗

Geometric consistency evaluation

We thank the reviewer for this valuable suggestion. We agree that more direct evaluations of garment fidelity and cross-view consistency are important. However, in 3DGS editing tasks, ground-truth edited 3D geometry and corresponding depth maps are generally unavailable. Therefore, geometric accuracy metrics that require ground-truth 3D surfaces or depth maps cannot be directly computed. We thus retain PSNR, SSIM, and LPIPS, which are also adopted in previous 3DGS editing studies, to ensure a fair comparison with existing methods.

To further address the concern regarding cross-view consistency, we additionally introduce MEt3R [Asim et al., CVPR 2025], a geometry-aware metric for evaluating multi-view consistency. Given a pair of views, MEt3R first employs DUSt3R to estimate dense 3D point maps. These point maps are then used to reproject DINO features, upsampled by FeatUp, into a common view, where bidirectional cosine similarity is computed over the overlapping regions. The final MEt3R score is defined as one minus the average bidirectional feature similarity, with lower scores indicating better cross-view consistency. Importantly, MEt3R does not require ground-truth 3D geometry, depth maps, or known camera poses. This metric has also been adopted to evaluate geometric consistency in 3D Gaussian head avatars in MVCHead [Chharia and De la Torre, CVPR 2026], further supporting its applicability to 3D Gaussian human representations.

We report MEt3R alongside the existing metrics in the target garment region comparison, as shown below.

\hline \text{Method} & \text{PSNR}\uparrow & \text{SSIM}\uparrow & \text{LPIPS}\downarrow & \text{MEt3R}\downarrow \\\\ \hline \text{GaussianEditor} & 19.4225 & 0.5290 & 0.2904 & 0.1264 \\\\ \text{DGE} & \mathbf{22.0834} & 0.5783 & 0.2085 & 0.1280 \\\\ \text{EditSplat} & 20.9711 & 0.5731 & 0.2157 & 0.1238 \\\\ \text{T3HG-Editor (Ours)} & 21.2198 & \mathbf{0.5796} & \mathbf{0.1738} & \mathbf{0.1187} \\\\ \hline \end{array}$$ As shown in the table, T3HG-Editor achieves the lowest MEt3R score of 0.1187, outperforming GaussianEditor (0.1264), DGE (0.1280), and EditSplat (0.1238). This result provides additional quantitative evidence that our method improves cross-view garment consistency, complementing the existing image-based evaluation metrics. References [1] Asim M, Wewer C, Wimmer T, et al. MET3R: Measuring Multi-View Consistency in Generated Images[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2025: 6034–6044. [2] Chharia A, De la Torre F. Multi-view Consistent 3D Gaussian Head Avatars 'without' Multi-view Generation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2026: 40163–40174.

I thank the authors for their response and clarifications. After considering the rebuttal, I will raise my rating to 3.

Reviewer_u4rYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes T3HG-Editor, a text-driven 3D human garment editing framework that leverages SMPL-X body priors to improve editing quality and cross-view consistency. The method combines SMPL-X-guided Gaussian seeding, vertex-based cross-view attention, and 2D/3D Gaussian pruning to preserve garment geometry and prevent unintended modifications. Experimental results demonstrate improvements over existing 3D Gaussian editing methods in text alignment, garment consistency, and non-target region preservation.

Critical strengths

The paper addresses a practical problem of text-driven 3D human garment editing. The proposed method demonstrates strong experiment results, outperforming existing 3D Gaussian editing approaches in terms of text alignment, garment consistency, and non-target region preservation. The qualitative comparisons and ablation studies further support the effectiveness of the proposed framework.

Critical weaknesses
  • W1. Limited Generalization to Diverse Human Subjects and Garments. The evaluation is limited to two human scenes and ten text prompts, making it difficult to assess the generalization of the proposed method to diverse human subjects and garments. In particular, the method is not evaluated under challenging conditions, such as complex poses, diverse body shapes, and occlusions. Furthermore, since SMPL-X primarily represents the human body surface, it remains unclear whether the proposed method can effectively handle loose-fitting garments, such as skirts or oversized coats, that deviate significantly from the underlying body geometry.
  • W2. Insufficient Robustness Analysis of SMPL-X Fitting Errors. The proposed method relies heavily on SMPL-X geometry across its main components, including Gaussian seeding, cross-view correspondence, and SDF-based pruning. Inaccurate SMPL-X fitting could lead to incorrectly initialized Gaussians, erroneous cross-view feature correspondences, or the unintended removal of valid Gaussians. However, the paper does not investigate how fitting errors affect editing performance or evaluate the robustness of the proposed method under inaccurate SMPL-X estimations, leaving its reliability in practical scenarios unclear.
  • W3. Insufficient Validation of SMPL-X Priors. Although leveraging SMPL-X priors is presented as a key contribution, the experiments do not sufficiently isolate and validate their effectiveness. The current ablation studies remove entire components, making it difficult to determine whether the observed improvements arise specifically from SMPL-X structural priors or from more general techniques such as Gaussian densification, multi-view attention, and pruning. Direct comparisons with alternative geometry-guided or correspondence-based approaches would help clarify the unique contribution of SMPL-X priors.
  • W4. Limited Evaluation of 3D Geometric Quality. Although the paper aims to improve garment shape adaptation and reduce geometric distortions, it lacks a direct quantitative evaluation of the resulting 3D geometry. The evaluation primarily relies on 2D rendering-based metrics, including CLIP similarity, PSNR, SSIM, and LPIPS, which do not directly measure the geometric accuracy of the edited garments. Additional analysis of geometric artifacts, floating Gaussians, and surface irregularities, particularly under garment expansion and shrinkage, would provide stronger evidence of the proposed method's effectiveness in preserving 3D geometric quality.
Questions for authors

I would appreciate it if the authors could address the concerns raised in the Weaknesses section, particularly regarding the limited generalization to diverse human subjects and garment types, the robustness to inaccurate SMPL-X fitting, the specific contribution of SMPL-X priors beyond general editing techniques, and the lack of direct evaluation of 3D geometric quality.

4 comments
AuthorsOpenReview ↗

Addition of four real-world scene editing experiments, clarification of SMPL-X-related issues, and introduction of the Editing Success Rate (ESR) metric

Thank you for the constructive comments. We have added experiments on four additional real-world scenes in Fig. 7 (p. 8) and introduced a new metric, the Editing Success Rate (ESR) in Section 4.3 (p. 9), with detailed evaluation criteria provided in the Appendix (p. 18–19).

(1) Limited Generalization to Diverse Human Subjects and Garments (W1).

First, we have added experiments on four real-world scenes in Fig. 7 of the rebuttal version, including two different environments and two human poses. These new scenes are captured under natural illumination, covering both relatively bright and dim lighting conditions. From Fig. 7, T3HG-Editor consistently achieves the best performance across different poses, environments, illumination conditions, and garment types.

Second, we sincerely appreciate the reviewer's suggestion regarding challenging human occlusions, which provides valuable guidance for extending our research. Existing image-guided 3DGS human virtual try-on methods, such as GaussianVTON [1] and GS-VTON [2], and general text-driven 3DGS editing methods, such as DGE [3] and EditSplat [4], have primarily conducted experiments involving standard human poses and relatively simple movements, but have not yet provided systematic evaluations under challenging human self-occlusion conditions (as also acknowledged in GS-VTON [2]). We acknowledge that handling such challenging cases deserves further investigation and consider this an important direction for future work.

Finally, our method can handle loose-fitting garments that deviate significantly from the underlying body geometry. The reported experiments cover a wide range of garment types and structural scales, from smaller, lightweight upper-body garments such as sleeveless tank tops to larger, more voluminous garments such as puffer jackets, and from shorts to full-length pants and dresses. We also evaluate suits and long garments with extended hems, such as trench coats and dresses. These experiments encompass diverse garment structures and seasonal clothing styles, providing further evidence of our method's effectiveness across different garment categories.

References for W1

[1] Chen et al., “GaussianVTON: 3D Human Virtual Try-ON via Multi-Stage Gaussian Splatting Editing with Image Prompting,” arXiv:2405.07472, 2024.

[2] Cao et al., “GS-VTON: Controllable 3D Virtual Try-on with Gaussian Splatting,” IJCV, 2026.

[3] Chen et al., “DGE: Direct Gaussian 3D Editing by Consistent Multi-view Editing,” ECCV, 2024.

[4] Lee et al., “EditSplat: Multi-View Fusion and Attention-Guided Optimization for View-Consistent 3D Scene Editing with 3D Gaussian Splatting,” CVPR, 2025.

(2) Insufficient Robustness Analysis of SMPL-X Fitting Errors (W2).

In our framework, the reconstructed 3DGS scene and the spatially aligned SMPL-X model serve as inputs, rather than components that we aim to improve. Similar to 3DGS reconstruction, SMPL-X fitting is a well-established upstream task supported by existing methods, including EasyMocap [5], SMPLify-X [6], AGORA [7], and PhysAvatar [8]. In particular, multi-view images, calibrated camera poses, and reconstructed 3D geometry provide additional constraints for SMPL-X fitting and spatial alignment. Furthermore, even if minor fitting errors exist, our cross-view attention and feature propagation mechanisms jointly process tokens associated with corresponding SMPL-X vertices and their neighboring features, making the effects of such minor errors negligible.

References for W2

[5] EasyMocap, “Make Human Motion Capture Easier,” GitHub, 2021.

[6] Pavlakos et al., “Expressive Body Capture: 3D Hands, Face, and Body from a Single Image,” CVPR, 2019.

[7] Patel et al., “AGORA: Avatars in Geography Optimized for Regression Analysis,” CVPR, 2021.

[8] Zheng et al., “PhysAvatar: Learning the Physics of Dressed 3D Avatars from Visual Observations,” ECCV, 2024.

(3) Insufficient Validation of SMPL-X Priors (W3).

We apologize for causing confusion and would like to clarify that SMPL-X is a mature parametric human model widely adopted in 3D human modeling and reconstruction. Adopting SMPL-X itself is not our contribution. Rather, our contributions lie in how SMPL-X is used, and we have ablated each component in the experiment section.

(4) Limited Evaluation of 3D Geometric Quality (W4).

We would like to clarify that 3DGS editing tasks do not have ground-truth edited results, making direct quantitative evaluation of the resulting 3D geometry infeasible. Previous 3DGS editing methods also evaluate with 2D rendering-based metrics (e.g., CLIP similarity, PSNR, SSIM, LPIPS), and we follow this established protocol for fair comparison. We also believe that developing dedicated geometric evaluation methods for 3DGS editing could be a valuable direction for future research.

AuthorsOpenReview ↗

Geometric consistency evaluation

We thank the reviewer for this valuable suggestion. We agree that more direct evaluations of garment fidelity and cross-view consistency are important. However, in 3DGS editing tasks, ground-truth edited 3D geometry and corresponding depth maps are generally unavailable. Therefore, geometric accuracy metrics that require ground-truth 3D surfaces or depth maps cannot be directly computed. We thus retain PSNR, SSIM, and LPIPS, which are also adopted in previous 3DGS editing studies, to ensure a fair comparison with existing methods.

To further address the concern regarding cross-view consistency, we additionally introduce MEt3R [Asim et al., CVPR 2025], a geometry-aware metric for evaluating multi-view consistency. Given a pair of views, MEt3R first employs DUSt3R to estimate dense 3D point maps. These point maps are then used to reproject DINO features, upsampled by FeatUp, into a common view, where bidirectional cosine similarity is computed over the overlapping regions. The final MEt3R score is defined as one minus the average bidirectional feature similarity, with lower scores indicating better cross-view consistency. Importantly, MEt3R does not require ground-truth 3D geometry, depth maps, or known camera poses. This metric has also been adopted to evaluate geometric consistency in 3D Gaussian head avatars in MVCHead [Chharia and De la Torre, CVPR 2026], further supporting its applicability to 3D Gaussian human representations.

We report MEt3R alongside the existing metrics in the target garment region comparison, as shown below.

\hline \text{Method} & \text{PSNR}\uparrow & \text{SSIM}\uparrow & \text{LPIPS}\downarrow & \text{MEt3R}\downarrow \\\\ \hline \text{GaussianEditor} & 19.4225 & 0.5290 & 0.2904 & 0.1264 \\\\ \text{DGE} & \mathbf{22.0834} & 0.5783 & 0.2085 & 0.1280 \\\\ \text{EditSplat} & 20.9711 & 0.5731 & 0.2157 & 0.1238 \\\\ \text{T3HG-Editor (Ours)} & 21.2198 & \mathbf{0.5796} & \mathbf{0.1738} & \mathbf{0.1187} \\\\ \hline \end{array}$$ As shown in the table, T3HG-Editor achieves the lowest MEt3R score of 0.1187, outperforming GaussianEditor (0.1264), DGE (0.1280), and EditSplat (0.1238). This result provides additional quantitative evidence that our method improves cross-view garment consistency, complementing the existing image-based evaluation metrics. References [1] Asim M, Wewer C, Wimmer T, et al. MET3R: Measuring Multi-View Consistency in Generated Images[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2025: 6034–6044. [2] Chharia A, De la Torre F. Multi-view Consistent 3D Gaussian Head Avatars 'without' Multi-view Generation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2026: 40163–40174.

Thank you to the authors for their responses and additional experiments. I appreciate the expanded evaluation on real-world scenes and the introduction of MEt3R, which partially address my concerns regarding generalization (W1) and geometric consistency (W4).

However, my concerns regarding SMPL-X fitting robustness (W2) and the validation of the proposed SMPL-X-specific contributions (W3) remain unresolved. In particular, I remain unconvinced that the proposed approach offers sufficient methodological novelty beyond existing techniques.

Given these remaining concerns, I will maintain my original rating of 2.

AuthorsOpenReview ↗

Further clarification on SMPL-X fitting robustness and methodological contributions

We sincerely thank the reviewer for revisiting our rebuttal and acknowledging the additional experiments and MEt3R evaluation. We appreciate the remaining concerns regarding SMPL-X fitting robustness (W2) and methodological novelty (W3), and would like to provide further clarification.

W2: Robustness to SMPL-X Fitting Errors. We appreciate the reviewer's concern and would like to clarify that although T3HG-Editor leverages SMPL-X structural priors, its editing process is not solely determined by the accuracy of SMPL-X fitting. Rather, SMPL-X provides approximate body geometry and cross-view correspondences, while the original 3DGS reconstruction, multi-view image supervision, and semantic constraints jointly guide the editing process.

Specifically, (1) Gaussian initialization: SMPL-X provides spatial guidance for seeding additional Gaussians, whose attributes are initialized from the original reconstructed 3DGS through KNN aggregation. More importantly, the fusion of pre- and post-edit semantic masks constrains and filters the editable Gaussians according to the actual garment regions, enabling reliable selection even when SMPL-X fitting is imperfect. This substantially reduces the dependence on precise SMPL-X alignment. (2) Cross-view consistency: Our method establishes correspondences through SMPL-X vertices while incorporating neighboring features, which helps mitigate local correspondence inaccuracies. (3) Gaussian pruning: Our pruning strategy does not rely solely on SMPL-X-based 3D geometric constraints, but more importantly combines them with complementary 2D semantic-mask constraints. In particular, 2D pruning effectively identifies and removes redundant Gaussians caused by garment shrinkage. As demonstrated in Fig. 5, Fig. 7, and Fig. 12, this joint pruning strategy effectively handles garment shrinkage while preserving plausible garment structures.

Furthermore, our experiments on independently reconstructed real-world scenes with different poses and illumination conditions (Fig. 7) demonstrate the practical applicability of this framework across different capture settings. We acknowledge that these results do not constitute a controlled fitting-error robustness evaluation. Nevertheless, the combination of reconstructed scene geometry, neighborhood-aware feature aggregation, and 2D/3D constraints reduces reliance on exact SMPL-X alignment. Moreover, SMPL-X fitting is a well-established research task, with numerous mature methods capable of achieving accurate fitting under suitable capture conditions. Therefore, obtaining reliable SMPL-X estimates is a tractable upstream task rather than a fundamental obstacle to the practical application of our framework.

W3: Specific Contributions of SMPL-X Priors. We would like to emphasize that our contribution is not the SMPL-X model itself or the independent use of existing editing operations, but the task-specific integration of SMPL-X structural priors into three complementary mechanisms for garment editing. Specifically, (1) normal-guided Gaussian seeding provides additional editable Gaussians for garment expansion beyond the original garment geometry; (2) explicit SMPL-X vertex correspondences establish anatomically aligned cross-view feature associations, rather than relying solely on general geometric or epipolar matching; and (3) SMPL-X-derived SDF constraints complement 2D semantic masks to suppress Gaussian drift while handling garment shrinkage.

Importantly, the ablation results in Fig. 8 and Table 3 demonstrate the effectiveness of these mechanisms, while the mask-guided comparisons in Fig. 13 show that existing methods still struggle with garment expansion even when target-region masks are provided. The sensitivity analysis in Fig. 9 further demonstrates the importance of adapting Gaussian seeding and geometric constraints to different garment structures. Although these experiments do not independently isolate SMPL-X from every possible alternative geometric prior, they provide evidence that our proposed body-prior-guided designs address challenges not adequately handled by general-purpose 3DGS editing pipelines.

We respectfully hope these clarifications and the additional experimental evidence help address the reviewer's remaining concerns. We sincerely appreciate the reviewer's constructive feedback and consideration.

Reviewer_bK3iOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses text-driven 3D human garment editing based on 3D Gaussian Splatting, and proposes T3HG-Editor, a three-stage framework leveraging SMPL-X body priors. The first stage obtains editable Gaussians by seeding Gaussians along SMPL-X surface normals and filtering them with fused pre-edit and post-edit masks. The second stage enforces garment consistency via cross-view attention and feature propagation, aligned by shared SMPL-X vertex correspondences. The third stage updates Gaussians while pruning overflowing ones using dual constraints: 2D semantic masks and an SDF distance field derived from SMPL-X. Experiments on two human subjects show that the method outperforms general-purpose 3DGS editing baselines in terms of editing quality, non-target region preservation, and cross-view consistency.

Critical strengths
  1. The paper targets a well-motivated practical problem: text-driven garment editing of 3D human assets, with clear application value in virtual try-on, digital content creation, and e-commerce.
  2. The three-stage pipeline is logically coherent and technically complete. Each module is designed to address a specific pain point (insufficient Gaussian coverage, cross-view inconsistency, non-target region contamination), and the overall workflow forms a closed engineering loop.
Critical weaknesses
  1. Experimental scale is too small to support generalizability claims. The evaluation is conducted on only 2 human subjects and 10 text prompts. With such a limited sample size, it is difficult to verify whether the method generalizes across different body shapes, poses, garment categories, and illumination conditions.
  2. Academic novelty is very limited. All core components are well-established techniques adapted to the garment editing scenario:
    • SMPL-X guided Gaussian initialization has been widely used in human 3DGS generation and reconstruction works (e.g., HumanGaussian, GaussianAvatar).
    • Cross-view attention and feature propagation for consistency follow similar ideas as general 3D editing methods like DGE and EditSplat, with only the correspondence anchor changed to SMPL-X vertices.
    • SDF-based geometric pruning is a standard operation in 3D editing and reconstruction. The work is essentially a systematic integration of existing techniques into a specific application scenario, with no fundamental methodological or conceptual breakthrough.
Questions for authors
  1. The experiments are conducted on only 2 human subjects. Why is the dataset so small? Have you tested the method on more diverse body shapes, poses, and garment types? How would you justify the generalizability of the method?
2 comments
AuthorsOpenReview ↗

Addition of four real-world garment editing experiments and clarification of our three main contributions

Thank you for your insightful comments.

W1: We have added experiments on four real-world scenes in Fig. 7 (p.8) of the rebuttal version, including two different environments and two human poses. These new scenes are captured under natural illumination, covering both relatively bright and dim lighting conditions. From Fig. 7, T3HG-Editor consistently achieves the best performance across different poses, environments, illumination conditions, and garment types.

Furthermore, our experiments cover diverse garment categories and geometric deformation scales, ranging from sleeveless tank tops to voluminous puffer jackets, and from shorts to full-length pants and dresses. We also evaluate suit jackets and long-hem garments such as trench coats, further demonstrating the effectiveness of our method across different garment structures.

W2: We would like to clarify that our work goes beyond a systematic integration, as each component involves task-specific designs. Garment editing involves large deformation (expansion or shrinkage), which existing pipelines cannot handle, so T3HG-Editor exploits SMPL-X priors across all three stages.

Our first contribution is obtaining editable Gaussians adapted to garment expansion, unlike HumanGaussian and GaussianAvatar, which primarily focus on human generation and reconstruction. HumanGaussian relies on SMPL-X surface-based Gaussian initialization for text-to-3D human generation, whereas our method explicitly seeds additional Gaussians along SMPL-X surface normals to accommodate garment expansion during editing. Our sensitivity analysis further shows that insufficient outward seeding leads to incomplete sleeves and garment hems. GaussianAvatar learns Gaussian offsets from a body template for animatable human reconstruction, while our method introduces additional editable Gaussians to support text-driven garment transformations in pre-reconstructed 3DGS scenes. Moreover, our mask-guided filtering localizes these editable Gaussians to the intended garment regions, avoiding unnecessary modifications to the original scene.

Our second contribution is cross-view attention and feature propagation with substantial task-specific improvements. (1) DGE performs global cross-view attention, potentially mixing unrelated features from backgrounds and garments, whereas our method introduces SMPL-X-guided attention within corresponding human-body regions to reduce such interference. (2) DGE employs epipolar-constrained feature matching, whereas our method leverages explicit SMPL-X vertex correspondences to reduce matching ambiguity. (3) DGE selects reference key views based on camera distance, while our method prioritizes them according to shared-vertex coverage, providing more anatomically relevant references. (4) During feature propagation, DGE fuses features from the two nearest key views using camera-distance-based weights, whereas our method aggregates features from four key views using adaptive weights derived from shared-vertex coverage.

Our third contribution is 2D and 3D joint pruning rather than conventional SDF pruning alone. Specifically, 2D semantic-mask constraints remove Gaussians extending beyond the target regions and, more importantly, identify and eliminate redundant Gaussians caused by garment shrinkage. Meanwhile, 3D SDF-based geometric constraints suppress spatially drifting Gaussians that may not be reliably detected from 2D projections alone. As shown in Fig. 5, Fig. 7 and Fig. 12, this handles garment shrinkage while preserving plausible structures.

AuthorsOpenReview ↗

Follow-up on our rebuttal and additional experiments

Dear Reviewer,

We sincerely thank you again for your constructive feedback. We have carefully addressed your concerns regarding experimental generalization and methodological novelty through substantial revisions to our manuscript.

Specifically, we added experiments on four real-world scenes with different environments, illumination conditions, human poses, and garment types (Fig. 7, p. 8). We also provided detailed clarifications of our three methodological contributions, highlighting their differences from existing approaches and supporting them with additional experimental evidence and sensitivity analyses.

We would greatly appreciate it if you could review our rebuttal and revised manuscript and share any further feedback, particularly regarding whether our additional experiments and clarifications have addressed your original concerns.

Thank you very much for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary and Key Revisions

Dear Area Chair,

We sincerely thank the reviewers and the Area Chair for their constructive feedback. During the rebuttal period, we carefully addressed all three reviewers' concerns through additional experiments, detailed responses, and manuscript revisions.

Reviewer bK3i has not yet provided follow-up feedback on our rebuttal. Reviewer u4rY acknowledged our additional experiments and MEt3R evaluation, and we provided further clarifications regarding their remaining concerns, for which we have not yet received additional feedback. Importantly, Reviewer CHjb acknowledged our rebuttal and raised their rating from 2 to 3 (Weak Acceptance).

We briefly summarize our key improvements below:

Reviewer bK3i (Generalization and Novelty). We added four real-world scenes covering different environments, illumination conditions, human poses, and garment types (Fig. 7, p. 8). We also clarified our three contributions: Gaussian seeding for garment expansion, SMPL-X-guided cross-view consistency, and joint 2D/3D pruning for garment shrinkage, supported by ablation and sensitivity analyses.

Reviewer u4rY (Generalization, SMPL-X Priors, and Geometric Evaluation). We added four real-world scenes covering different environments, illumination conditions, human poses, and garment types (Fig. 7, p. 8). We also provided further clarifications regarding SMPL-X fitting robustness and methodological contributions, and introduced the geometry-aware MEt3R metric, achieving the best score of 0.1187.

Reviewer CHjb (Editing Fidelity and Cross-view Consistency). We introduced the Editing Success Rate (ESR) and MEt3R, providing additional evidence of editing effectiveness and cross-view consistency.

We respectfully hope the Area Chair will consider our additional evidence, detailed rebuttal responses, and revised manuscript in the final assessment.

Thank you for your time and consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.