acceptodds
Withdrawn from ICLR 2027

CoStyler: Indoor Furniture Styling by Counterfactual Reasoning in a Hypergraph Field

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Fixed-layout indoor furniture styling requires selecting assets that form a coherent room while preserving furniture categories, positions, orientations, and scales. Existing approaches typically retrieve each asset independently or rely on static local relations, making them prone to shape, material, and color conflicts after scene composition. We introduce CoStyler, a scene-level structured selection framework built on a dynamic hypergraph style field. It uses a frozen multimodal large language model to extract style priors, maintains candidate distributions for furniture slots, and captures higher-order dependencies through style-conditioned hyperedge activation and weighting. Counterfactual style preference learning evaluates candidates as local substitutions using Mahalanobis energies to measure contextual compatibility. Training alternates between optimizing the style field and candidate logits; inference freezes the model and iteratively updates only room-specific candidate logits to resolve cross-slot conflicts. Experiments on 3D-FRONT show that CoStyler outperforms object- and scene-level retrieval baselines in furniture retrieval and scene-level style coherence.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Indoor Scene Synthesis, 3D Asset Retrieval

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_sg4dOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies joint furniture selection under a fixed room layout. Instead of retrieving each furniture item independently, CoStyler first retrieves a candidate set for every furniture slot and then uses a dynamic hypergraph to model higher-order stylistic relationships among slots. Candidate compatibility is evaluated by locally substituting an item and measuring the resulting change in scene energy. At test time, the model parameters are frozen while the candidate distributions for the current room are iteratively optimized to obtain a complete furniture assignment. The method is evaluated on 3D-FRONT. The results show improvements over the compared methods in exact furniture retrieval, automatic scene-level evaluation, and human evaluation. The ablations indicate that iterative test-time optimization and the Mahalanobis energy are the main contributors to performance.

Critical strengths
  1. The scene-level joint-selection problem is practically relevant. Furniture items retrieved independently may each satisfy the textual request while becoming inconsistent in color, material, shape, or scale when composed into a room. Formulating the task as room-level joint optimization is better aligned with the actual furniture-styling objective than independent object retrieval.

  2. The dynamic hypergraph and counterfactual compatibility mechanism are well matched to the problem. The hypergraph can model higher-order relationships among multiple furniture items, while counterfactual substitution allows the same candidate to receive different compatibility scores in different room contexts. The experiments and ablations suggest that joint optimization substantially improves scene-level coherence.

Critical weaknesses
  1. The target-style descriptions may contain information derived from the ground-truth furniture. The target text is generated by a vision-language model after observing the original fully furnished room, which already contains the furniture that the system is expected to recover. Even though the prompt requests a room-level description, the resulting color, material, and form information still originates from the target scene. The evaluation may therefore be closer to reconstructing furniture from a visual summary of the original room than to styling a room from an independently written user request. This could overestimate performance in realistic applications.

  2. Unclear train–test asset overlap and insufficient scene-level baselines make the large improvement difficult to interpret. The dataset is split by room, but the paper does not clearly state whether furniture asset IDs are disjoint between training and testing. If identical assets or similar furniture combinations occur in both splits, the model may learn asset-level co-occurrence patterns. In addition, most baselines perform independent object retrieval, whereas CoStyler uses approximately 200 steps of scene-level test-time optimization. Without an asset-disjoint evaluation and stronger joint-selection baselines, the increase in R@1 from 22.8 to 79.1 cannot be attributed conclusively to the proposed method.

Questions for authors
  1. The target-style description is generated from the original fully furnished room. How much information about the specific target furniture may be retained in this description? Would CoStyler remain effective when using independently written descriptions created without observing the target furniture?

  2. What fraction of furniture asset IDs is shared between the training and test sets? Can the authors provide results on an asset-disjoint split that also controls for near-duplicate assets and repeated furniture combinations?

  3. What is the recall of the ground-truth asset within the initial top-15 candidate set? Please report GT Recall@15 to clarify the upper bound imposed by candidate retrieval.

  4. Can the authors compare CoStyler against stronger scene-level alternatives, such as pairwise compatibility reranking, a standard graph model, beam search, or multimodal-model-based scene reranking?

  5. What are the runtime, memory consumption, and computational cost of approximately 200 test-time optimization steps per room? How do these costs scale with the number of furniture slots?

6.Since multiple furniture combinations may be equally appropriate, how do the authors view the discrepancy between exact R@1 and the actual furniture-styling objective? Could the evaluation place greater emphasis on non-exact scene-level quality measures?

Reviewer_2RjgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

CoStyler addresses fixed-layout indoor furniture styling: given a style description and furniture slots with fixed categories, positions, orientations, and scales, it jointly selects assets from a library. Frozen Qwen3-VL models generate style priors and retrieve candidates; a style-conditioned hypergraph propagates scene context over candidate distributions, and Mahalanobis energies score scene compatibility and local substitutions. Training alternates field and candidate-logit updates; inference freezes the field and optimizes room-specific logits. On 3D-FRONT, CoStyler reports strong reference-asset recovery (R@1 79.1% vs 22.8% at init; baseline comparisons are subject to verification, see Weaknesses).

Critical strengths
  1. A practical task with a coherent joint-selection formulation. Maintaining candidate distributions and revising them against evolving scene context sensibly avoids committing prematurely to independent retrieval results, and the components form a connected inference loop.

  2. Strong reported gains and human evidence, pending verification of the baseline numbers. Reference R@1 reaches 79.1% (vs 22.8% at initialization), and CoStyler's own blinded human ratings approach those of the original reference rooms (e.g., human style coherence 4.61 vs 4.65 for the reference scenes). The reported margins over baselines depend on baseline numbers whose provenance requires verification (Weakness 1).

Critical weaknesses
  1. The baseline results reported in Table 1 appear to have been taken from MetaFind's original paper rather than produced by the described evaluation. The four baseline Final R@1 values — 33.9 (ULIP), 35.1 (OpenShape), 36.3 (Uni3D), 44.5 (MetaFind) — are exactly equal to the R@1 values in MetaFind (Pan et al., NeurIPS 2025), Table 1, under the text+point-cloud query column on Objaverse-LVIS (R@1/R@5: 33.9/58, 35.1/61.4, 36.3/63.6, 44.5/71.3). These evaluations were conducted under different conditions: MetaFind reports open multimodal retrieval on Objaverse-LVIS and scene-quality ratings of I-Design scenes by GPT-4o and five annotators, whereas CoStyler reports slot-level exact-match R@1 on 3D-FRONT with shared top-15 candidate pools and scene ratings by GPT-4o and ten experts. Independent evaluations under different datasets and protocols would not produce identical two-decimal values. The numbers therefore appear to have been transcribed from the original paper rather than generated by the described experiments.

  2. The protocol does not separate reference recovery from transferable styling. Style descriptions are extracted from images of the original furnished rooms, reference configurations are used as training positives, and the main metric measures exact recovery of the original assets. Recovering a professionally arranged composition can itself yield high human ratings, so the human study supports output quality but not the ability to coordinate alternative assets under independently specified styles. I do not interpret this as leakage or memorization; it bounds what the current protocol can conclude. A direct test would apply multiple independently authored styles to the same layouts, or exclude original and near-duplicate assets from the candidate pools, and evaluate style adherence and inter-asset compatibility separately.

  3. The necessity of the proposed architecture is not established. Table 3 never removes the hypergraph: the only learned row without the Mahalanobis metric (hypergraph + iteration, Euclidean energy) improves over the retrieval prior by just 0.9 points (23.7 vs 22.8), so the evidence does not rule out that the hypergraph is largely inert. All learned variants retain the hypergraph, and there is no matched comparison against a pairwise graph, a Transformer, or a static hypergraph, nor an isolation of style-conditioned gating or counterfactual scoring. Without such controls, the contribution is best supported as an effective task-specific system rather than a demonstrated advance in higher-order reasoning.

Questions for authors
  1. Candidate-energy training. Which loss trains ϕc\phi_c and the predictors of (μi,σi)(\mu_i,\sigma_i)? Eq. (3) ranks only scene energies, while Eq. (4) uses candidate energies with the field frozen; absent explicit parameter sharing, these heads appear to receive no gradient. Please specify the complete implemented objective, parameter sharing, and how refined training-room logits re-enter network training.

  2. Evaluation protocol and candidate construction. The shared top-15 pool is built by the paper's own frozen retriever, so object-level baselines are evaluated only within it; please report each baseline with its own candidate pool or justify the shared protocol. Are training and test rooms split at the residence level, and are duplicate or near-duplicate layouts removed? Please report asset overlap across rooms, whether original assets can appear in candidate pools, and how slots whose reference asset is absent from the pool are handled in R@1.

  3. Architecture necessity (assessment-changing). Please compare against a graph/Transformer contextual model with matched features, energy heads, and inference budget, and against a VLM that directly selects assets from rendered candidates. Together with a test of independently specified styles or non-reference compositions (Weakness 2), these would directly substantiate the paper's central claims.

Reviewer_f4UvOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents CoStyler, a framework for fixed-layout indoor furniture styling that aims to select stylistically coherent furniture assets while preserving the original scene layout. The method employs a frozen multimodal large language model to extract room- and slot-level style priors and maintains probability distributions over furniture candidates. A dynamic hypergraph models higher-order relationships among furniture slots, while counterfactual preference learning evaluates candidate substitutions using scene- and candidate-level Mahalanobis energies. The framework further adopts alternating optimization during training and iterative test-time optimization to refine furniture selections. Experiments on 3D-FRONT report substantial improvements in furniture retrieval accuracy and scene-level style coherence.

Critical strengths
  1. The paper addresses a meaningful and practical problem. Fixed-layout furniture styling requires balancing individual asset relevance with overall scene coherence while preserving spatial constraints. This setting is relevant to indoor scene design and 3D content creation.

  2. The proposed framework provides a coherent approach to scene-aware furniture selection. By combining distributional candidate representations, style-conditioned hypergraph reasoning, and counterfactual candidate evaluation, the method allows furniture selections to be refined according to the evolving scene context rather than relying on independent retrieval. These designs are well motivated by the specific challenges of the task.

  3. The paper provides relatively comprehensive experimental evaluations. The experiments include comparisons with object-level and scene-aware retrieval methods, ablation studies, convergence analysis, and qualitative results. The inclusion of both GPT-based and human evaluations provides additional evidence regarding the perceptual quality of the generated furniture compositions.

Critical weaknesses
  1. The reported baseline results raise serious concerns about the reliability of the quantitative evaluation. Table 1 reports R@1 values of 33.9, 35.1, 36.3, and 44.5 for ULIP, OpenShape, Uni3D, and MetaFind, respectively. These four values exactly match those reported in Table 1 of the original MetaFind paper under the text-plus-point-cloud retrieval setting on Objaverse-LVIS. This is particularly concerning because CoStyler presents these results as obtained on 3D-FRONT under a fixed-layout furniture retrieval protocol, which involves a different dataset and evaluation setting. The paper also states that MetaFind was independently reimplemented because its source code was unavailable. The authors should clarify the source of these baseline results and whether they were independently evaluated under the stated protocol.

  2. The ablation study does not sufficiently isolate the contributions of the dynamic hypergraph and counterfactual candidate evaluation. Table 3 demonstrates substantial improvements from iterative optimization and Mahalanobis energy. However, all variants using iterative optimization also retain the hypergraph, making it difficult to determine the independent contribution of higher-order reasoning. Furthermore, although counterfactual candidate evaluation is presented as a central contribution, the current ablations mainly compare Mahalanobis and Euclidean energies rather than directly examining the effect of counterfactual candidate scoring itself. Removing the hypergraph while retaining iterative optimization, or removing the counterfactual candidate-energy term while preserving the remaining framework, would provide more direct evidence for the effectiveness of these core design choices.

  3. The training mechanism of the candidate-level Mahalanobis energy is unclear. Section 3.4 defines both scene-level and candidate-level energies. However, Eq. (3), which trains the style-field parameters, only includes scene-level energy ranking and hyperedge sparsity regularization. The candidate-level energy is subsequently used in Eq. (4), where the style-field parameters are frozen and only candidate logits are optimized. Since the main objective of CoStyler is to identify which furniture candidates best match the surrounding objects and target style, it is unclear how the model learns to distinguish compatible and incompatible candidates at the individual furniture level under this training setup. The authors should clarify how the candidate-level energy is learned and how it contributes to selecting stylistically coherent furniture.

Questions for authors

Seen above.

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

Related papers

Discussion (0)

Sign in to comment.