acceptodds
Withdrawn from ICLR 2027

Roomer: Reflective Object-Grounded Model Editing and Repair for 3D Indoor Layout Synthesis

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing indoor layout generators produce globally plausible layouts yet may retain local violations such as collisions, out-of-bounds placements, obstructed openings, and blocked circulation. Most prior work focuses on full-scene synthesis or scene-level optimization, with limited support for identifying responsible objects and locally repairing affected regions. We present Roomer, a reflective repair framework that casts these violations as sparse, object-grounded repair problems. Roomer encodes layouts as "RoState" and uses "RoReview" to bind measured violations to implicated objects. A geometry-conditioned vision-language model planner proposes a structured local edit, while a deterministic solver validates it and generates a finite set of candidate edits when needed. Each candidate is committed only if full-scene verification confirms that it resolves the target violation without new hard violations or broken protected constraints. We train the planner on Roomer-CC, a controlled-corruption dataset that pairs faulty layouts with object-grounded violation evidence and known-feasible inverse StatePatches. Since existing benchmarks rarely assess whether physically valid layouts are usable, we introduce Roomer-Eval to assess distributional quality, physical validity, and practical usability. Experiments show that Roomer repairs residual violations while preserving valid regions, improves physical validity and usability, and transfers across external generators.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Indoor Layout Synthesis, Object-Grounded Repair

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_BTUPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Roomer, a framework that combines object-grounded geometric diagnosis, a geometry-conditioned vision-language planner, deterministic search, and verification with rollback to locally repair generated 3D indoor layouts while seeking to preserve already-valid content. The authors also introduce the Roomer-CC training dataset and Roomer-Eval evaluation protocol, with experiments demonstrating reduced geometric violations and improved satisfaction of predefined spatial usability rules on held-out 3D-FRONT rooms and outputs from four external generators.

Critical strengths

S1. Clear problem formulation and a controlled local repair mechanism. The paper addresses specific violations in existing layouts by combining object-grounded diagnosis, geometry-conditioned planning, candidate search, and verification with rollback. Each edit targets an identifiable object and violation, and is checked for newly introduced hard-rule violations before being committed. This design provides a practical way to correct local errors while preserving already-valid parts of the layout. S2. A dedicated training dataset and reasonably comprehensive evaluation provide direct evidence of repair effectiveness. The authors construct a dataset of 67,550 paired repair examples and evaluate geometric validity and practical usability, supplemented by professional judgments. In particular, the paired before-and-after evaluations on outputs from four external generators retain all test scenes, providing convincing evidence that Roomer improves existing layouts within the evaluated setting.

Critical weaknesses

W1. The absence of a direct comparison with closely related recent work leaves the incremental contribution insufficiently established. SceneReVis already supports visual feedback, multi-turn object-level editing, and the repair of existing flawed layouts, making it closely related to Roomer in both task and overall workflow. Although Roomer cites this work, it provides neither a detailed methodological comparison nor a quantitative evaluation against it, and does not explicitly explain its exclusion. Consequently, the current experiments do not sufficiently establish the additional benefits of Roomer’s object-grounded diagnosis, geometry-conditioning branch, and verification mechanism over this close alternative. W2. The ablation intended to demonstrate the advantage of local editing over full-scene rewriting is confounded. Beyond changing the output scope, the Full-Scene JSON Rewrite variant removes the acceptance conditions that prohibit new hard-rule violations and preserve protected satisfied relations, and does not use the corresponding deterministic candidate fallback search. Its weaker performance may therefore reflect the removal of safeguards and search, rather than the use of full-scene rewriting itself. This comparison does not isolate the contribution of the local StatePatch representation.

Questions for authors

Q1. Could the authors compare Roomer with more recent, closely related baselines, such as SceneReVis? Q2. Given the potential generalization capabilities of the general-purpose Qwen2.5-VL backbone, could the authors evaluate Roomer beyond 3D-FRONT?

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

Roomer addresses residual local faults in generated 3D indoor layouts as small, object-level repair problems. The system converts a completed layout into RoState, an object-addressable representation, and builds RoReview, which ties each measured violation to the objects and geometric evidence involved. A geometry-conditioned vision-language planner then proposes a structured local edit, called a StatePatch, and a deterministic solver checks that proposal and, when needed, produces a finite set of alternative candidates. A candidate is committed only if a full-scene floor-plane check confirms that the target violation is resolved without new hard-rule violations or broken protected relations. To train the planner, the authors build Roomer-CC, a controlled-corruption dataset of 67,550 paired repair examples with object-grounded evidence and known-feasible inverse patches. They also introduce Roomer-Eval, which combines standard distributional and physical-validity metrics with five reproducible rule families for practical usability. Validation uses a frozen cohort of 1,100 held-out 3D-FRONT rooms, density-controlled comparisons, removal of individual components, and frozen transfer to outputs from four external generators.

Critical strengths
  1. The paper contributes a repair Roomer-CC dataset, filling a gap in scenario design/tuning.

  2. The paper is clearly structured and features a robust experimental design.

  3. The experiments demonstrate the effectiveness of the framework's repair mechanism: the diagnostics are correlated with specific furniture instances and their measurement relationships, and are validated before repair is performed, which is a reasonable approach to correcting residual faults while maintaining the integrity of valid content. Tong Shi ablation experiments directly support the learned planner and validation gates: removing validation and rolling back dramatically increases the number of new hard errors, and the entire loop achieves a very high, strictly safe repair rate.

Critical weaknesses
  1. The notation, formulas and intermediate representations make it difficult to trace what was detected, what the model proposed, what the solver changed, and why the edit was accepted. These steps are central to the method, so the main text should connect them more clearly.
Questions for authors
  1. Could you demonstrate a complete repair process in the main text, from identifying the problem to proposing, implementing, accepting, or reversing the changes? It would be even better if the author could explain the key symbols in the examples and clarify the respective functions of the model and the deterministic solver.

  2. Do these rules check the orientation relationships between furniture (e.g., a sofa facing a TV), or only the space and circulation? If such relationships are not covered, how does this limit the method's improvement in practical applications?

  3. What range of furniture counts and scene densities does the test set cover? From what I've observed in the results presentation, the number of assets in the scene content is relatively small, so I'm concerned about how the repair success rate will change as the scene contains more objects or more interaction issues.

Reviewer_qELDOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces Roomer, an object-grounded framework that locally repairs collisions, out-of-bounds placements, obstructed openings, and usability violations in generated 3D indoor layouts. Roomer combines RoState and RoReview representations, a geometry-conditioned VLM planner, deterministic StatePatch search, and verification-gated commitment, and introduces Roomer-CC for training and Roomer-Eval for evaluating physical validity and practical usability. Experiments show that Roomer improves layout validity and usability across multiple external generators while largely preserving unaffected scene regions.

Critical strengths
  • The paper addresses local violations in 3D indoor layouts while preserving valid regions.

  • The framework combines VLM planning, deterministic search, and verification for reliable local repair.

  • Extensive comparisons, ablations, and cross-generator evaluations demonstrate its effectiveness and generalizability.

Critical weaknesses
  1. The paper assumes that residual violations are typically sparse and localized, but lacks quantitative evidence across different generators. Moreover, identifying objects involved in a violation does not necessarily determine which object should be edited. An analysis of violation distributions and cases requiring multiple-object edits would better support this assumption.

  2. The proposed "Practical" metric relies on geometric proxies rather than directly assessing furniture accessibility and usability. For example, bedside clearance does not guarantee access from the entrance, while the living-room angle rule overlooks sofa orientation. More direct task-oriented evaluation would strengthen the claims.

  3. The necessity of fine-tuning a 7B VLM remains unclear given the relatively constrained room types evaluated. Analyzing performance across varying scene complexities would help establish when such a large model is beneficial.

  4. The paper lacks comparisons with constraint-based optimization methods, such as those explored in Holodeck and Infinigen Indoors (CVPR 2024). Adapting these approaches to the same initial layouts and comparing repair quality, content preservation, and computational cost would better justify the proposed framework.

Questions for authors
  1. What is the distribution of violations and affected objects per scene, and how often are multiple-object edits necessary?

  2. Can the authors evaluate practical usability using task-oriented measures of furniture accessibility and interaction?

  3. How does repair performance vary with scene complexity, particularly in rooms containing three to five distinct furniture groups?

  4. Can the authors compare against constraint-based optimization baselines adapted from methods such as Holodeck, AnyHome, Infinigen Indoors, or SceneLCM under the same initial layouts and repair constraints?

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

Related papers

Discussion (0)

Sign in to comment.