acceptodds
Under review as a conference paper at ICLR 2027

GDAvatar: Geometry-Guided Dynamic Gaussian Head Avatars from Monocular Videos

Abstract

Recent Gaussian head avatars achieve remarkable rendering quality and real-time performance, but coarse canonical geometry and limited motion information hinder faithful reconstruction of expression-dependent appearance details. In this paper, we propose GDAvatar, a geometry-guided dynamic Gaussian head avatar framework for monocular videos. We first optimize the canonical mesh using a vertex-bound Gaussian representation to obtain personalized geometry. Building upon the optimized geometry, we then employ a triangle-bound Gaussian representation to model dynamic Gaussian attribute offsets by jointly encoding canonical and dynamic features. An attention fusion module adaptively balances these features, enabling more accurate modeling of local non-rigid deformations. Furthermore, we construct a continuous editable Gaussian map from the learned representation, enabling flexible and high-quality texture editing. Extensive experiments on multiple datasets demonstrate that GDAvatar achieves superior rendering quality and animation fidelity compared with state-of-the-art methods, while naturally supporting high-quality texture editing.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, 3D Head Avatars

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_xAyJOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
2
Presentation
3(+1)
Soundness
3(+1)
Read review
Paper summary

This paper proposes GDAvatar, a monocular Gaussian head avatar framework that progressively allocates Gaussian degrees of freedom. It first optimizes personalized canonical geometry using constrained vertex-bound Gaussians, then models dynamic appearance using triangle-bound Gaussians and a geometry-conditioned UV offset branch, and finally constructs a continuous UV representation for appearance editing.

Critical strengths

The motivation and overall pipeline are clear and intuitive. The method achieves consistently strong reconstruction results across the evaluated datasets. The ablations cover most major components and generally support the proposed design.

Critical weaknesses

The methodological novelty appears somewhat limited. The main contribution is the progressive restriction and release of Gaussian degrees of freedom, which is intuitive and effective, but appears closer to a carefully designed staged optimization strategy than a fundamentally new representation or modeling principle.

The central geometry–appearance separation claim is only indirectly validated. Since the evaluated datasets do not provide ground-truth geometry, improved rendering quality does not necessarily demonstrate that the recovered canonical geometry is more accurate or that geometry and dynamic appearance are better disentangled.

The editing capability is mainly demonstrated qualitatively, with limited quantitative evaluation of editing quality, consistency across expressions/poses, or preservation of unedited regions.

Questions for authors

How can the authors verify that progressive optimization actually improves canonical geometry, rather than simply finding a different geometry–appearance decomposition that achieves better rendering quality?

To what extent should progressive DoF allocation be viewed as a general modeling principle rather than a task-specific staged optimization strategy? Is the benefit primarily due to the representation design or simply the optimization schedule?

Why is the canonical geometry completely frozen in Stage B? Would allowing small learning-rate geometry refinement after releasing the appearance DoFs provide a better trade-off?

To my knowledge, some prior works have also explored structured or topology-aware constraints on Gaussian attributes (e.g., TAvatar[1]); would the authors consider citing it?

Can the authors provide quantitative evaluation of the editing capability, especially consistency across different expressions or poses and preservation of regions that are not edited?

[1]Tavatar: Topology-Aware Gaussian Attribute Derivation for Animatable Human Avatars, cvpr, 2026

3 comments
AuthorsOpenReview ↗

Response to Reviewer xAyJ (Part 1)

Thank you for your valuable comments. We have addressed the concerns through methodological novelty clarification, quantitative geometry evaluation, geometry freezing ablation and editing consistency clarification.

Q1: Explain on Methodological Novelty.

Ans: Our contribution goes beyond a conventional staged optimization strategy by explicitly separating identity-specific geometry from expression-dependent appearance variations through progressive Gaussian DoF allocation. Unlike conventional staged optimization, which primarily changes when parameters are optimized, our method controls what the representation can model at each stage. Stage A restricts Gaussian DoFs to recover personalized canonical geometry, while Stage B releases dynamic appearance DoFs on the fixed geometric basis.

As shown in Table 4, Joint Optimization, which optimizes geometry and appearance simultaneously without progressive DoF allocation, yields inferior results, demonstrating the benefit of explicitly separating their optimization. Furthermore, GeoRefine, which continues updating geometry in Stage B, also degrades performance, suggesting that geometry changes interfere with learning stable UV-space Gaussian attribute offsets. These findings support progressive DoF allocation as a representation-level modeling principle rather than merely an optimization schedule.

Table4: Ablation study of Progressive DoF Allocation on the INSTA dataset.

\small \begin{array}{lccc} \hline \text{Method} & \text{PSNR}\uparrow & \text{SSIM}\uparrow & \text{LPIPS}\downarrow \\\\ \hline \text{w/ Joint Optimization} & 30.476 & 0.951 & 0.058 \\\\ \text{w/ GeoRefine in Stage B} & 30.892 & 0.956 & 0.052 \\\\ \hline \mathbf{Full\\ Model} & \mathbf{31.431} & \mathbf{0.960} & \mathbf{0.043} \\\\ \hline \end{array} }$$ **Q2. Verify Geometry Quality.** **Ans:** To verify that progressive optimization improves canonical geometry rather than merely rendering quality, we introduce additional geometric evaluations. On INSTA, which lacks ground-truth 3D geometry, we evaluate reconstructed mesh silhouettes using Silhouette IoU and Boundary Chamfer Distance. Our method (0.851↑ / 23.914 px↓) outperforms Template and VHAP as shown in **Table 2** (Page 8). More importantly, on H3DS with ground-truth 3D meshes, our method achieves the best L1-Chamfer Distance and Normal Consistency (0.836×10⁻²↓ / 0.809↑) across 10 subjects, as reported in **Figure 11** and **Table 5** (Page 17) of the supplementary material. These results demonstrate improved geometric accuracy beyond rendering quality. **Table2: Quantitative geometry evaluation on the INSTA dataset.** $${ \small \begin{array}{lcc} \hline \text{Method} & \text{Silhouette IoU}\uparrow & \text{Boundary Chamfer (px)}\downarrow \\\\ \hline \text{Template} & 0.785 & 26.937 \\\\ \text{VHAP} & 0.819 & 24.572 \\\\ \hline \mathbf{Ours} & \mathbf{0.851} & \mathbf{23.914} \\\\ \hline \end{array} }$$ **Table5: Quantitative geometry evaluation on the H3DS dataset.** $${ \small \begin{array}{lcc} \hline \text{Method} & \text{L1-Chamfer}\downarrow & \text{N. C.}\uparrow \\\\ \hline \text{Template} & 0.979\times 10^{-2} & 0.776 \\\\ \text{VHAP} & 0.940\times 10^{-2} & 0.780 \\\\ \hline \mathbf{Ours} & \mathbf{0.836\times 10^{-2}} & \mathbf{0.809} \\\\ \hline \end{array} }$$ **Q3: Geometry Freezing in Stage B.** **Ans:** Freezing canonical geometry in Stage B provides a stable geometric basis for learning UV-space Gaussian attribute offsets. Further geometry refinement would continuously change this basis, interfering with dynamic appearance learning. To validate this choice, we evaluate GeoRefine, which allows continued geometry refinement in Stage B. As shown in **Table 4**, **GeoRefine** (30.892 / 0.956 / 0.052) underperforms our full model (31.431 / 0.960 / 0.043) in PSNR / SSIM / LPIPS. These results indicate that maintaining fixed geometry during Stage B is more beneficial than continued geometry refinement.
AuthorsOpenReview ↗

Response to Reviewer xAyJ (Part 2)

Q4: Related Work TAvatar.

Ans: Thank you for pointing out this relevant work. TAvatar analytically derives Gaussian attributes from mesh topology to maintain surface consistency under novel poses. In contrast, our method progressively separates identity-specific geometry from expression-dependent appearance and learns geometry-conditioned Gaussian attribute offsets for fine-grained facial dynamics. We have added TAvatar to the Related Work section.

Q5: Evaluation of Editing Capability.

Ans: Thank you for your suggestion. In Stage C (Editable Gaussian Map Construction), Gaussian attributes are represented by continuous UV attribute maps. Since all frames share the same UV coordinate space, sticker editing maintains consistent facial correspondence across different expressions and poses. Moreover, editing modifies only the selected UV region, leaving attributes outside this region unchanged. As stickers occupy only a small region, the reconstruction quality of unedited regions is expected to remain comparable to the full-face results. In Table 8 of the original manuscript (Page 18), we demonstrated the face reconstruction quality across different expressions and poses compared to GT images; the results indicate that this editable representation maintains high-quality reconstruction.

Thanks for the clarification! The authors have addressed all my concerns. I will maintain my Weak Accept rating.

Reviewer_AikyOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents GDAvatar, a geometry-guided framework for reconstructing dynamic Gaussian head avatars from monocular videos. The key idea is to progressively allocate the degrees of freedom of Gaussian primitives to disentangle identity-specific geometry from expression-dependent appearance variations. The method first optimizes personalized canonical geometry using constrained vertex-bound Gaussians, and then employs triangle-bound Gaussians together with a geometry-conditioned UV Offset Branch to model dynamic appearance. Finally, the learned representation is converted into a continuous UV Gaussian map to enable appearance editing. Experiments on three monocular head-avatar datasets demonstrate strong reconstruction and animation performance.

Critical strengths
  1. The progressive Gaussian DoF allocation strategy is well motivated and empirically effective. The method first restricts Gaussian DoFs to encourage reconstruction errors to be explained by personalized geometry, and then releases additional DoFs for dynamic appearance modeling. I think this strategy makes sense and follows a natural geometry-first principle that has also been explored in earlier avatar reconstruction works, while the progressive allocation of Gaussian DoFs provides an effective realization of this idea. The joint-optimization ablation further demonstrates the benefit of this strategy.

  2. The method achieves strong quantitative results across INSTA, PointAvatar, and NerFace, together with convincing self-reenactment and cross-identity reenactment results. The ablations cover the major components, including geometry optimization, dynamic features, attention fusion, the UV Offset Branch, and progressive optimization.

Critical weaknesses
  1. The overall pipeline feels somewhat like a hybrid of ideas from several existing avatar reconstruction methods. Many individual components, such as personalized geometry optimization, mesh-bound Gaussians, learned dynamic offsets, and UV-space appearance modeling, are relatively familiar and closely related to prior works such as PointAvatar, FateAvatar, and other recent Gaussian avatar methods.
Questions for authors
  1. Are expression, pose, and camera parameters refined in both Stage A and Stage B? A fixed-tracking ablation may help isolate the effect of progressive DoF allocation.
  2. How does the progressive DoF strategy differ from the geometry-to-appearance modeling in prior works such as MeGA and Gaussian Head Avatar?
  3. The concurrent MGAvatar appears to follow a closely related geometry-to-appearance strategy. What are the main differences from GDAvatar?
  4. Have the authors considered evaluating on data with ground-truth 3D geometry? This may help directly validate the claimed improvement in personalized geometry.
3 comments
AuthorsOpenReview ↗

Response to Reviewer Aiky (Part 1)

Thank you for your valuable comments. We have addressed the concerns through novelty clarification, quantitative geometry evaluation, tracking refinement ablation, comparison with related methods, and an ethics statement.

Q1: Explain on Novelty.

Ans: Our work is not simply a combination of existing techniques, but addresses the ambiguity between identity-specific geometry and expression-dependent variations through progressive Gaussian DoF allocation. PointAvatar represents geometry using deformable points without explicit mesh topology, while FateAvatar optimizes mesh-bound Gaussian attributes without directly refining the underlying geometry. In contrast, our method first recovers personalized canonical geometry through constrained Gaussian DoFs, then uses this geometry as a stable basis to guide expression-dependent appearance modeling through geometry-conditioned dynamic Gaussian offsets. This separation of geometric refinement and dynamic appearance learning distinguishes our approach from prior methods.

Q2: Evaluation on datasets with GT 3D geometry.

Ans: Since the INSTA dataset lacks ground-truth 3D geometry, we first evaluate 2D geometric consistency by comparing rendered masks of reconstructed meshes with ground-truth image masks on the test set, using Silhouette IoU and Boundary Chamfer Distance. As newly reported in Table 2 (Page 8), our method outperforms both the FLAME template and the VHAP geometric model, indicating better alignment between the reconstructed geometry and the observed head silhouettes.

To directly evaluate 3D geometric accuracy, we further conduct experiments on the H3DS dataset[1] with ground-truth 3D geometry. Across 10 subjects, our method achieves the best L1-Chamfer Distance and Normal Consistency, outperforming Template and VHAP. We add these results in Figure 11 and Table 5 (Page 17) of the supplementary material, demonstrating more accurate personalized 3D geometry reconstruction. Together, these evaluations validate our method in terms of both 2D silhouette consistency and 3D geometric accuracy.

[1] Ramon E, Triginer G, Escur J, et al. H3d-net: Few-shot high-fidelity 3d head reconstruction[C]. ICCV 2021.

Table2: Quantitative geometry evaluation on the INSTA dataset.

\small \begin{array}{lcc} \hline \text{Method} & \text{Silhouette IoU}\uparrow & \text{Boundary Chamfer (px)}\downarrow \\\\ \hline \text{Template} & 0.785 & 26.937 \\\\ \text{VHAP} & 0.819 & 24.572 \\\\ \hline \mathbf{Ours} & \mathbf{0.851} & \mathbf{23.914} \\\\ \hline \end{array} }$$ **Table5: Quantitative geometry evaluation on the H3DS dataset.** $${ \small \begin{array}{lcc} \hline \text{Method} & \text{L1-Chamfer}\downarrow & \text{N. C.}\uparrow \\\\ \hline \text{Template} & 0.979\times 10^{-2} & 0.776 \\\\ \text{VHAP} & 0.940\times 10^{-2} & 0.780 \\\\ \hline \mathbf{Ours} & \mathbf{0.836\times 10^{-2}} & \mathbf{0.809} \\\\ \hline \end{array} }$$ **Q3: Refinement on expression, pose, and camera parameters.** **Ans:** Expression, pose, and camera parameters are refined in both stages, as clarified in Sec. 3.2 (Page 5). To assess the impact of tracking refinement, we added a **"w/ Fixed Tracking"** ablation in **Table 3** (Page 9), keeping the initial tracking parameters fixed throughout both stages. The results show that this leads to a decrease in PSNR. In comparison, **"w/o Geometry Optimization"**, which skips the constrained-DoF geometry stage and directly optimizes full-DoF triangle-bound Gaussians, results in a larger PSNR drop. This suggests that progressively establishing personalized geometry before releasing additional Gaussian DoFs contributes more to reconstruction quality than tracking refinement. **Table 3: Ablation study on the INSTA dataset.** (Part of experiments.) $${ \small \begin{array}{lccc} \hline \text{Method} & \text{PSNR}\uparrow & \text{SSIM}\uparrow & \text{LPIPS}\downarrow \\\\ \hline \text{w/o Geometry Optimization} & 30.544 & 0.953 & 0.051 \\\\ \text{w/~~ Fixed Tracking} & 30.954 & 0.957 & 0.046 \\\\ \hline \mathbf{Full\\ Model} & \mathbf{31.431} & \mathbf{0.960} & \mathbf{0.043} \\\\ \hline \end{array} }$$
AuthorsOpenReview ↗

Response to Reviewer Aiky (Part 2)

Q4: Differences with MeGA, Gaussian Head Avatar, and MGAvatar.

Ans: MeGA refines geometry via UV displacements, while Gaussian Head Avatar uses SDF/DMTet initialization followed by joint Gaussian and deformation optimization. Neither explicitly separates geometric refinement from flexible appearance optimization, potentially allowing appearance to compensate for geometric errors. MGAvatar adopts a geometry-to-appearance strategy but directly mixes geometric and latent features without distinguishing identity-specific and expression-dependent information. In contrast, GDAvatar progressively allocates Gaussian DoFs to establish stable personalized geometry. More importantly, it explicitly uses this geometry to guide dynamic appearance modeling by adaptively fusing canonical and dynamic features, enabling spatially adaptive identity preservation and expression-dependent deformation modeling. Table 4 shows that Joint Optimization (simultaneous geometry–appearance learning) and GeoRefine (continued geometry refinement during appearance learning) underperform our full model supporting the importance of progressive optimization and stable geometry.

Table4: Ablation study of Progressive DoF Allocation on the INSTA dataset.

\small \begin{array}{lccc} \hline \text{Method} & \text{PSNR}\uparrow & \text{SSIM}\uparrow & \text{LPIPS}\downarrow \\\\ \hline \text{w/ Joint Optimization} & 30.476 & 0.951 & 0.058 \\\\ \text{w/ GeoRefine in Stage B} & 30.892 & 0.956 & 0.052 \\\\ \hline \mathbf{Full\\ Model} & \mathbf{31.431} & \mathbf{0.960} & \mathbf{0.043} \\\\ \hline \end{array} }$$ **Q5: Ethics statement.** **Ans:** We have added an Ethics Statement (Page 10), clarifying that facial data are sourced from public research datasets and addressing informed consent, privacy protection, and risks of unauthorized impersonation and misuse.

Thanks for the clarification. The authors have adequately addressed my concerns, and the paper is in good shape overall. I will maintain my Weak Accept rating.

Reviewer_vyTjOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes GDAvatar, a monocular Gaussian head avatar framework that progressively allocates Gaussian degrees of freedom: Stage A optimizes a personalized canonical mesh withGaussians with constrained DoFs. Stage B freezes geometry and models expression-dependent attribute offsets with triangle-bound Gaussians plus a geometry-conditioned UV Offset Branch with attention fusion. Stage C distills base attributes into a continuous UV map. Experiments on INSTA, PointAvatar, and NerFace report improved PSNR/SSIM/LPIPS over several recent baselines

Critical strengths

The progressive allocation of Gaussian DoFs (vertex-bound constrained DoF for canonical geometry, then triangle-bound full DoF for dynamic appearance) is well-motivated to mitigate representation ambiguity and appears to stabilize geometric learning under monocular supervision. In addition, the editable UV representation broadens applicability to practical editing tasks and shows consistent appearance under pose/expression changes.

Critical weaknesses

The authors claim that GDAvatar achieves superior rendering quality and animation fidelity. However, Table 1 only reports image reconstruction metrics (PSNR, SSIM, and LPIPS), while cross-identity reenactment is evaluated qualitatively without quantitative metrics. Given that the paper's central motivation is to reduce the entanglement between identity-specific geometry and expression-dependent variations, cross-identity reenactment is a particularly important setting for evaluating this claim. Quantitative evaluations of expression transfer accuracy and identity preservation would provide stronger evidence for the claimed improvement in animation fidelity.

Questions for authors
  1. Effect of mesh subdivision: The proposed method subdivides the FLAME mesh from approximately 5K to 20K vertices, which increases its geometric representation capacity. How much of the improvement comes from this subdivision rather than the proposed geometry optimization? Would the baselines also benefit from applying the same subdivision?
  2. Progressive DoF release: In Section 3.2, does "progressively release" simply refer to the transition from Stage A to Stage B, or is there an additional schedule for gradually releasing Gaussian degrees of freedom within Stage B? The current description is somewhat unclear.
2 comments
AuthorsOpenReview ↗

Response to Reviewer vyTj

Thank you for your valuable comments. We have addressed the concerns through quantitative cross-identity reenactment evaluation, mesh subdivision ablation, and clarification of progressive Gaussian DoF release.

Q1:Quantitative evaluation of cross-identity reenactment.

Ans: Following common practices in 3D head avatar reconstruction (e.g., GaussianAvatars and STAvatar), our original evaluation focused on reconstruction metrics and qualitative reenactment results. We have now added quantitative cross-identity reenactment evaluations using CSIM (Identity Similarity), AED (Average Expression Distance), and APD (Average Pose Distance), following UIKA [1]. As shown in Supplementary Table 7 (page 18), our method achieves the best results across all three metrics, outperforming FateAvatar and RGBAvatar. These quantitative evaluations support the effectiveness of our geometry-aware modeling in improving expression transfer accuracy and identity preservation.

[1] Wu Z, Zhou B, Hu L, et al. UIKA: Fast Universal Head Avatar from Pose-Free Images[C]. CVPR 2026.

Table 7: Quantitative comparison of cross-reenactment performance on the INSTA, PointAvatar, and NerFace datasets.

\small \begin{array}{lccc} \hline \text{Method} & \text{CSIM}\uparrow & \text{AED}\downarrow & \text{APD}\downarrow \\\\ \hline \text{FlashAvatar (CVPR'24)} & 0.828 & 0.104 & 0.066 \\\\ \text{MonoAvatar (SIGGRAPH'24)} & 0.776 & 0.108 & 0.068 \\\\ \text{GaussianAvatars (CVPR'24)} & 0.739 & 0.109 & 0.071 \\\\ \text{RGBAvatar (CVPR'25)} & 0.830 & 0.103 & 0.060 \\\\ \text{FateAvatar (CVPR'25)} & 0.835 & 0.102 & 0.059 \\\\ \text{STAvatar (CVPR'26)} & 0.808 & 0.105 & 0.064 \\\\ \hline \mathbf{Ours} & \mathbf{0.838} & \mathbf{0.100} & \mathbf{0.056} \\\\ \hline \end{array} }$$ **Q2:Effect of mesh subdivision on reconstruction quality.** **Ans:** We have added an ablation study (**"w/o Subdivision"**) in **Table 3** (Page 9). Our full model achieves the best performance. Removing subdivision moderately degrades performance, while removing geometry optimization causes a larger drop, indicating that geometry optimization contributes more significantly to reconstruction quality. In our previous experiments, we applied subdivided mesh to GaussianAvatars; we found that if the mesh is not optimized during the reconstruction process, subdivision yields negligible improvements to rendering quality. **Table3: Ablation study on the INSTA dataset.** (Part of experiments.) $${ \small \begin{array}{lccc} \hline \text{Method} & \text{PSNR}\uparrow & \text{SSIM}\uparrow & \text{LPIPS}\downarrow \\\\ \hline \text{w/o Geometry Optimization} & 30.544 & 0.953 & 0.051 \\\\ \text{w/o Subdivision} & 30.795 & 0.955 & 0.047 \\\\ \hline \mathbf{Full\\ Model} & \mathbf{31.431} & \mathbf{0.960} & \mathbf{0.043} \\\\ \hline \end{array} }$$ **Q3:Clarification of progressive Gaussian DoF release.** **Ans:** The progressive DoF release refers to the transition from Stage A to Stage B, rather than a gradual schedule within Stage B. In Stage A, vertex-bound Gaussians with constrained DoFs are used to optimize personalized canonical geometry. In Stage B, the optimized geometry is frozen, and Gaussians switch to triangle-bound representations with full DoFs and adaptive densification. We have clarified this in Section 3.2 (Page 4).

The authors have addressed my concerns. I will increase my score from 2 to 3.

Comments on the paper

2 comments
Area_Chair_cXE9OpenReview ↗

Invitation to Participate in the Discussion

The authors have provided their responses to the reviewers' comments. We would appreciate it if the reviewers could review the responses and share any further thoughts during the discussion period.

AuthorsOpenReview ↗

Rebuttal Summary

We sincerely thank all reviewers for their constructive feedback. The reviewers raised important questions concerning cross-identity reenactment, geometric accuracy, mesh subdivision, tracking refinement, progressive Gaussian DoF allocation, and editing capability. We have addressed these concerns through additional experiments, clarifications, and manuscript revisions.

The main revisions are summarized as follows:

  • Cross-identity evaluation (App. A.3.3, Tab. 7): Added quantitative reenactment metrics (CSIM, AED, APD) on the original datasets to assess identity preservation and expression transfer accuracy.
  • Geometry evaluation (Sec. 4.3, Tab. 2; App. A.3.2, Fig. 11 and Tab. 5): Added 2D silhouette evaluation on INSTA and 3D geometry evaluation on H3DS with ground-truth meshes.
  • Mesh subdivision and tracking (Sec. 4.4, Tab. 3): Added ablations showing that progressive DoF allocation contributes more to reconstruction quality than mesh subdivision or tracking refinement.
  • Progressive DoF analysis (Sec. 4.4, Tab. 4): Added ablations on joint geometry–appearance optimization (Joint Optimization) and continued geometry refinement during appearance learning (GeoRefine), validating the benefits of progressive geometry-to-appearance optimization.
  • Clarifications (Sec. 2.2; App. Tab.8): Clarified key differences from prior works, editing capability, and ethical considerations.

We are grateful for the reviewers’ valuable feedback, which has substantially improved our work. Detailed responses to all comments are provided in the corresponding reviewer discussions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.