acceptodds
Withdrawn from ICLR 2027

SLatHarmony: Locality-Guided Harmonization for Local 3D Editing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Local 3D editing requires faithfully realizing the intended changes within the target region, preserving non-edited regions, and maintaining coherent shape and texture across the entire asset. Existing methods typically mitigate non-target changes through constrained generation or inversion-based editing. In contrast, we leverage the spatial locality of the structured latent (SLat) representation in TRELLIS.2 to combine source and edited SLat fields through regional composition, thereby preserving source content outside the target region. However, achieving coherent shape and texture integration between target and non-edited regions remains challenging. We therefore propose SLatHarmony, a framework that couples regional SLat composition with cross-field harmonization. Specifically, a shape SLat seamer predicts locally decodable transition features using context from the source and edited sides to bridge geometric discrepancies near the interface, while a texture SLat harmonizer combines channel-wise field alignment with boundary-aware target-region correction to reduce appearance mismatch. Experiments on Edit3D-Bench show that SLatHarmony substantially improves the preservation of non-edited geometry and appearance while maintaining competitive semantic alignment with the edit instruction.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D generative modeling, 3D editing, structured latent representation, cross-field harmonization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_iB1nOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes SLatHarmony, a framework for local 3D editing that aims to preserve non-edited regions while maintaining geometric and texture consistency. The authors observe that existing methods often introduce unintended changes to non-edited regions, while directly combining source and edited SLat features can lead to geometric discontinuities and texture inconsistencies at the boundaries. To address these issues, the authors propose a shape SLat seamer to refine geometric transitions and a texture SLat harmonizer to reduce appearance mismatches between the original and edited regions.

The method is evaluated on Edit3D-Bench against three existing TRELLIS-based editing methods, along with several component-level ablation studies. The experimental results show improved preservation of non-edited geometry and appearance while maintaining competitive semantic alignment with the editing instructions. Overall, the paper presents a relatively straightforward and practical approach to addressing boundary consistency issues in local 3D editing.

Critical strengths

The authors propose a reasonable and straightforward method for addressing boundary inconsistencies in local 3D editing. The experimental results show promising improvements in preserving non-edited regions while maintaining competitive editing quality.

Critical weaknesses
  1. The experimental evaluation is not sufficiently comprehensive. The paper only compares against three existing methods, and the ablation studies are not thorough enough. In particular, the authors do not provide a comparison with direct SLat composition without the proposed modules, or an ablation that uses only the channel-wise texture shift without the learned residual correction. These experiments are important to understand whether the proposed modules are actually necessary and how much improvement they bring over simpler alternatives. Moreover, SLatHarmony uses a 512-resolution TRELLIS.2 configuration, while the baselines use 1024-resolution configurations, which raises additional concerns about the fairness of the comparisons. Overall, I feel that the current experimental evidence is not sufficiently convincing.
  2. The methodological novelty is relatively limited. While the proposed approach is reasonable, I find it somewhat engineering-oriented. The method mainly combines latent feature composition with relatively standard feature completion and alignment techniques. Although the results are promising, I am not convinced that these components offer sufficient technical novelty for a main conference like ICLR. In my opinion, the work may be more suitable for a workshop.
Questions for authors
  1. Regarding the experimental evaluation: I feel that the current experimental evaluation is not sufficiently comprehensive. The number of baselines is relatively limited, and the ablation studies are not thorough enough to fully justify the proposed design. More importantly, SLatHarmony uses the 512-resolution TRELLIS.2 configuration with a 32³ sparse grid, while the compared baselines use 1024-resolution configurations with 64³ sparse grids. This makes the experimental settings inconsistent across methods. Could the authors explain how this difference affects the fairness of the comparisons and provide more comprehensive experimental evidence to support their claims?
  2. Regarding the methodological novelty: Could the authors further clarify what the key technical novelty of the proposed method is, beyond combining existing feature completion and alignment techniques? I feel that the current design is relatively straightforward and engineering-oriented. It would be helpful to understand why these specific modules are necessary and what makes the proposed approach fundamentally different from simpler alternatives or existing SLat-based editing methods.
Reviewer_v2LuOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a framework for local 3D editing. The method uses regional composition of SLat features to preserve non-edited regions, and introduces a Shape SLat Seamer to handle interface inconsistencies between source and target latents. A Texture SLat Harmonizer is used to reduce appearance differences. Experiments on Edit3D-Bench show improved preservation of non-edited regions.

Critical strengths
  1. The paper addresses a practical problem in local 3D editing.
  2. The method shows good results in preserving non-edited regions.
Critical weaknesses
  1. In Section 4.1, the authors analyze spatial locality. However, regional composition has already been explored in previous works (Nano3D (Ye et al., ICLR 2026), VoxHammer (Li et al., 3DV 2026)). The proposed mask-based composition of sparse structures and SLat features does not seem to offer much novelty. It is also unclear how this analysis helps determine the boundary width or harmonization scope.
  2. The Shape SLat Seamer seems like a masked latent inpainting module. TRELLIS has already shown local editing using RePaint (Lugmayr et al., CVPR 2022). The corresponding technical contribution is limited compared with existing inpainting-based editing methods.
  3. During training, the Texture SLat Harmonizer uses two texture SLat fields which are sampled under the same condition with different sampling processes (as I understand it, the difference comes from using different random seeds). However, inference uses source and edited fields from different conditions. It is unclear whether the differences learned during training can generalize to this setting.
  4. For the main comparison results, Nano3D and VoxHammer were originally based on TRELLIS, while the proposed method uses TRELLIS.2. The authors do not clarify whether these baselines were adapted to TRELLIS.2. Differences in the backbone may affect the performance because TRELLIS.2 has better generation performance than TRELLIS.
Questions for authors
  1. The definition of response R(d)R(d) in Section 4.1 is unclear. What metrics are used to measure the shape and texture responses? Could the authors give the mathematical formulation of R(d)R(d)?
  2. Table 2 only evaluates the full Texture Harmonizer against hard splicing. Could the authors provide results of mean-shift-only and residual-only baselines?
  3. I think the global channel-wise mean shift may not capture spatially varying feature differences. For example, if the differences are mainly near the editing boundary, averaging over the entire non-edited region may weaken the shift. Have the authors examined how the feature differences change with distance to the boundary?
  4. Do the authors test the method's robustness to 3D mask errors? It would be useful to test its robustness when the mask includes some non-edited regions or when the mask boundary is shifted.
  5. The Shape Seamer uses a fixed one-voxel boundary width during inference. Is it possible to provide an ablation with different widths?
Reviewer_ZUDUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies local 3D editing with TRELLIS.2 structured latents, where the goal is to change a masked region while keeping the rest of the asset unchanged. It argues that shape and texture latents are spatially local, so source latents can be copied outside the edit mask and edited latents used inside it. To fix mismatches at the interface, it adds a learned shape module that predicts boundary latents from both sides and a texture module that applies a channel-wise shift plus a learned residual inside the mask. Experiments on Edit3D-Bench report better preservation of non-edited regions than prior TRELLIS-based editors, with comparable edit alignment.

Critical strengths
  • Clear and simple design. The locality study in Sec 4.1 and Fig 1 motivates composing latents directly, and Eq 1-2 make feature provenance explicit -- source support and features are kept exactly outside the mask, and only a one-voxel band plus the masked texture are modified. This makes the method easy to analyze and to plug in after other structure editors.
  • The texture module is a useful, practical part. It is trained on same-source texture pairs and doesn't need paired edit data. Under the same shape treatment, it lowers FID from 248.39 to 218.12 and regional LPIPS from 0.060 to 0.053, and the visible color seams in Fig 4 go away.
Critical weaknesses

W1. Tab 1 does not establish a gain over the state of the art.

  • Baselines disagree with published numbers. On the same Edit3D-Bench edits and masks, VoxHammer reports masked SSIM 0.994, LPIPS 0.027, FID 23.05, CLIP-T 0.287; here it gets 0.694, 0.112, 236.99, 0.226. The proposed method's 0.901 SSIM is below VoxHammer's own figure. The paper uses its own protocol but never discusses the gap.
  • Missing baseline. Easy3E [1] is TRELLIS-based and takes the same inputs (3D mask plus one edited view). It is cited in Sec 2.2, and the supplementary code even has an Easy3E path, yet it is absent from Tab 1.
  • Preservation is partly by construction. Eq 1 copies source support and latents outside the ground-truth mask, so the clean Fig 3 voxel plots are guaranteed. If the regional metrics use the TRELLIS.2 recon as the ref instead of the original mesh, the gain is inflated. This is not stated, abd the evaluation scripts are not in the supplement, and the CD scale is ambiguous: under the "x10^-3" label, 0.022 means about 2x10^-5, whereas VoxHammer reports unscaled CDs of 0.012-0.047 on the same benchmark.
  • Weak statistics. The results come from a single run with no variance and no user study. FID is computed on about 300 renders, and the CLIP-T range across methods is 0.225-0.234.

W2. The ablation does not support the learned shape module.

  • Composition does most of the work. The weakest Tab 2 variant already reaches CD 0.023 and SSIM 0.888, compared with 0.055-0.069 and 0.670-0.694 for the baselines.
  • Small shape effects. Replacing interpolation with the shape module changes CD from 0.023 to 0.022, leaves LPIPS at 0.273, and raises FID from 243.90 to 248.39. The support for it is mostly qualitative.
  • Training does not match the task. The shape module learns to refill a hidden band inside one coherent shape field (Eq 4), where both sides agree and simply continuing either side works. At inference the two sides come from independent generations and can disagree in surface position or thickness, which is the case the module is meant for. Sec 5.4 acknowledges this field-identity shift but does not measure it. The texture module, by contrast, is trained on independently generated pairs (Sec 4.4), so a similar setup seems feasible for shape. This mismatch may explain why the shape module is close to interpolation in Tab 2.
  • Missing controls. Tab 2 lacks four rows:
    • Shape hard splicing: copy source latents outside the mask and edited latents inside, with no boundary treatment. Tab 2 applies hard splicing only to texture.
    • Shift-only texture: App A.2 calls the shift "a strong parameter-free alignment baseline," but it is not reported alone.
    • Poisson blending [2]: a training-free blend of the texture latents on the voxel graph.
    • Shape module without the image input: its DINOv3 conditioning appears only in Appendix, not in Eq 3

W3. Pipeline description is incomplete.

  • Fig 2 omits the 3D mask, which defines Eq 1 and the scope of every module. The structure editor is described only as "the sparse-structure generation procedure of Vinedresser3D" (Sec 5.1), but the released code defaults to VoxHammer's.
  • Generated voxels outside the mask are discarded, so edits cannot exceed the annotated volume. The paper does not discuss this limit.

References

[1] Easy3E: Feed-Forward 3D Asset Editing via Rectified Voxel Flow, Hu et al., CVPR 2026

[2] Poisson Image Editing, Perez et al., SIGGRAPH 2003

Questions for authors
  1. Please report Tab 1 under the original Edit3D-Bench protocol used by VoxHammer, or explain why your VoxHammer numbers differ so much.
  2. Are the regional CD/LPIPS/SSIM computed against the original benchmark mesh or against a TRELLIS.2 reconstruction of it? What exactly does "CD (x10^-3)" mean (squared or not, normalization, number of points)? Please include the evaluation scripts.
  3. Please add Easy3E [1] to Tab 1 under the same mask, target image, and views.
  4. Which structure editor produced the Tab 1 results: the Vinedresser3D procedure stated in Sec 5.1 or the VoxHammer default in the code? How often does the generated structure extend beyond the mask and get truncated by Eq 1?
  5. Please add Tab 2 rows for hard splicing of shape latents with no boundary-band treatment, shift-only texture alignment, and the shape module without image conditioning. Please also compare the texture module with Poisson blending [2] of the texture latents on the 6-neighbor voxel graph: keep the edited field's neighbor differences inside the mask and fix the source values at its boundary. This is a training-free way to get a spatially varying offset (the Eq 5 shift is its constant special case), so it directly tests whether the learned residual is needed.
  6. Sec 5.1 says the baselines use "1024-resolution configurations with 64^3 sparse grids," but VoxHammer, Nano3D, and Vinedresser3D are all built on the original TRELLIS. Its latent grid is 64^3, mesh output is 256^3, so "1024" fits only TRELLIS.2. Were the baselines ported to TRELLIS.2-1024, or run on TRELLIS as released? If ported, why not run them at 512 to match SLatHarmony, so that resolution is not mixed with the method in Tab 1?
  7. How does the shape module behave when the two sides truly disagree? For example, train or test it on two independently sampled shape fields on the same sparse structure, as is done for texture, and report a metric restricted to the seam region (such as normal discontinuity or cracks near the interface). None of the current metrics isolates the seam.
  8. Does Tab 1 use all 300 instances, including the PartObjaverse-Tiny objects? Please report variance over seeds.
  9. Suggestion: since the method is presented as a module that runs after any upstream editor, applying it to the outputs of VoxHammer or Vinedresser3D, and with predicted rather than annotated masks (for example, from PartField), would show how general it is.

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

Related papers

Discussion (0)

Sign in to comment.