acceptodds
Under review as a conference paper at ICLR 2027

The Geometry of Consistent Intervention in Layered 3D Worlds

Abstract

Object-level edits to 3D worlds can break support and containment relations, while propagating an edit through the full dependency graph can move objects unnecessarily. We formulate translational intervention consistency in an object-position space where structural relations define a dependency-consistent feasible set. For a fixed support assignment, this set is a convex polyhedron induced by support equalities, support-surface inequalities, and containment constraints. We characterize how local interventions leave this feasible set and relate constraint violation to distance from feasibility. This geometry separates dependency from propagation: dependencies specify where an intervention may have consequences, while the intervention direction and local constraint geometry determine which relations actually require response. Consistency restoration is then formulated as a unique minimum-displacement projection within the appropriate feasible stratum, with stability guarantees under a fixed dependency structure. Experiments on reconstructed, RGB-D, synthetic, and generated 3D worlds evaluate prediction, propagation, repair, and consistency diagnosis. On InteriorGS, local geometry predicts intervention-induced inconsistency with AUROC 0.877, compared with 0.694 for graph fan-out. On a separate 340-edit benchmark, intervention-dependent propagation moves 59% fewer objects than full dependency closure while maintaining consistency in 334/340 edits. Independent physical rollouts, cross-dataset evaluations, and generated-world experiments further examine the relation between structural consistency and physical plausibility. Beyond post-hoc editing, the resulting feasible-state geometry provides an explicit consistency representation that can support verification and learning-time constraints in interactive and generative world models.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
3D world models, structured scene editing, intervention consistency, feasible-state geometry, dependency propagation, constraint projection, 3D scene generation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_FZsDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a consistent object editing framework in layered 3D worlds through a geometric formulation. Rather than local transformation rules, the paper adopt projection based approach to minimise undesired object movement while maintaining structural dependencies. I find the problem interesting and overall experiments are broad.

Critical strengths

Strengths:

  1. The geometric formulation is clear.
  2. The experimental evaluation spans across multiple datasets, baseline and robustness experiments.
Critical weaknesses

Weaknesses:

  1. In Table 2, the proposed method moved fewer objects, which is positive. However, physical displacement and hanging object metrics remains worse compared to the baseline. I understand there might be a trade off but a clear justification is necessary if reduction of undesired movements are preferable over other two.
  2. The paper lacks qualitative evidences, I tried to look into the supplementary too but could not find relevant evidences. Therefore, lack of visual evidence massively undermines the proposed methodology. The authors deliberately reported 1023 detected support relations and six object level inpainting interventions. However, exact benefit of the method remains unclear visually.
  3. Lack of detailed literature study.
Questions for authors

Questions for rebuttal:

  1. Weakness 1, please provide the justification in detail.
  2. Important: Please provide qualitative evidences showcasing original, edited, and repaired results.
  3. Line 324, there is a "??" in the eq. reference.
Reviewer_UNL9OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper formulates translational edit consistency through a dependency-consistent feasible set defined by support and containment relations. It characterizes intervention-induced constraint violation and proposes minimal-intervention propagation and practical projection solvers for consistency restoration. Experiments evaluate 142 translation, deletion, and container-constrained editing cases on InteriorGS, with additional evaluations on ARKitScenes, synthetic deep-support scenes, and HunyuanWorld-generated scenes.

Critical strengths
  • The paper distinguishes the dependency closure from the objects that must respond to an intervention. Using local constraint geometry to determine propagation is a useful approach to reducing unnecessary object movement.

  • Within a fixed stratum, the framework connects intervention-induced constraint violation, distance from feasibility, and the minimum-displacement consistent response.

Critical weaknesses
  • There is no visualized example in the whole submission, including the supplementary material. Regarding the paper's scope, providing visualized output is extremely important.

  • The “exact propagation set” claim is supported only by an edge-local criterion under fixed-assignment, unique-support assumptions. Theorem 3 establishes a minimum-displacement consistent response, not global ℓ0\ell_0-minimality. The introduction also incorrectly describes dist⁡(w,F)\operatorname{dist}(w,F) as piecewise linear, whereas this property applies to the aggregate constraint residual RFR_F.

  • ScanEdit[1] studies hierarchical scene editing with support-surface constraints and scene subgraph optimization. Edit-As-Act[2] formulates scene editing as goal-regressive planning that preserves the existing scene while performing necessary actions. I recommend that the authors add these works to the related work section.

  • The propagation-set F1 of 1.00 is an internal-consistency check because the geometric predictor and projection solver use the same center-of-gravity criterion. In the independent PyBullet rollout, the proposed method has a higher hanging count than follow (upward closure), 0.59 versus 0.06. Confidence intervals for improvements over the baselines are also missing from the main comparisons.

  • There are severe concerns in the submitted supplementary materials. (1) The provided projection solver fails on its own example and omits support-surface inequalities and grounding equalities. (2) The training code uses synthetic data instead of InteriorGS leave-one-scene-out repair and contains a broadcasting error in the supervised MSE. (3) The same 68-edit benchmark has conflicting results, including a physics-settling zero-residual rate of 20% in the paper versus 94% in the supplementary table. (4) supp.pdf seems not a final version; Did the authors check their supplementary material?

[1]: El Amine Boudjoghra, Mohamed, Ivan Laptev, and Angela Dai. "Scanedit: Hierarchically-guided functional 3d scan editing." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2025.

[2]: Noh, Seongrae, et al. “Edit-As-Act: Goal-Regressive Planning for Open-Vocabulary 3D Indoor Scene Editing.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026, pp. 19864–19873.

Questions for authors

See weaknesses.

Reviewer_jTdyOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper formulates 3D scene editing as a geometric constrained optimization problem, selectively propagates object movements based on violated support and containment constraints, and restores structural consistency through minimum-displacement projection, thereby reducing movements caused by dependency graph propagation.

Critical strengths
  1. Theoretical Analysis. The paper provides theoretical characterizations of the feasible-state geometry, intervention-induced constraint violations, and the existence and uniqueness of minimum-displacement solutions under fixed support assignments. These results offer a mathematical foundation for the proposed approach.
Critical weaknesses
  1. Lack of Visualizations and Qualitative Results. The paper lacks figures and qualitative results, which makes the proposed geometric intervention mechanism and its practical benefits difficult to assess intuitively. The authors should include illustrative diagrams and before-and-after 3D scene comparisons against baseline methods to better demonstrate the effectiveness of their approach.
  2. Missing Dataset Citation and Details. The paper extensively evaluates on InteriorGS but fails to cite its original source. The authors should provide the appropriate dataset reference and clarify the selection criteria for the 10 scenes used in their experiments to improve reproducibility.
Questions for authors

Please refer to the Weaknesses section.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.