Source-linked AI summary
StyleForge: Indoor Furniture Styling by Counterfactual Reasoning in a Hypergraph Field
Lingwei Dang, Shishuo Shang, Pan Liu, Jiajia Cheng, Ziyan Qiu, Zhenhao Zhang, Yufei Zhu, Shenghui Huang, Qingxin Xiao, Yun Hao, Juntong Li, Qingyao Wu
TL;DR
Fixed-layout furniture styling must produce a coherent room while preserving prescribed categories and geometry, but independent retrieval can create cross-furniture conflicts. StyleForge uses dynamic hypergraph reasoning and counterfactual refinement to select assets jointly, achieving state-of-the-art retrieval and improving over MetaFind by 34.6 R@1 points on 3D-FRONT.
Problem
Fixed-layout furniture styling must bridge object-level semantic relevance and scene-level aesthetic coherence while preserving each slot’s prescribed category, position, orientation, and scale.
Method
StyleForge models candidate distributions with a style-conditioned dynamic hypergraph field, evaluates contextual substitutions using Mahalanobis energies, and iteratively refines room-specific choices at inference.
Results
34.6 R@1 points: StyleForge improves over MetaFind, the strongest scene-aware baseline, while achieving state-of-the-art furniture retrieval and scene-level style coherence on 3D-FRONT.
Takeaways & Limitations
The results support jointly refining furniture selections because higher-order context can propagate corrections across slots as scene complexity increases.
Takeaways & Limitations
StyleForge depends on the initial retriever’s candidate recall because assets omitted from the initial top-K set cannot be reconsidered during scene-level optimization.
Abstract
from arXiv · showhide
Fixed-layout indoor furniture styling requires selecting assets that form a coherent room without changing the prescribed furniture categories, positions, orientations, or 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 StyleForge, a scene-level structured selection framework built on a dynamic hypergraph style field. A frozen multimodal large language model extracts structured style priors from an open-ended style request and the fixed layout, while StyleForge maintains a learnable candidate distribution for each furniture slot. Conditioned on the target style, the dynamic hypergraph style field adaptively activates and weights layout-induced hyperedges to capture higher-order dependencies among furniture. Counterfactual style preference learning then treats each candidate as a local substitution in the current style field and evaluates its contextual compatibility using Mahalanobis energies. Training alternates between optimizing the style field and the candidate logits. At inference, the model remains frozen and test-time training updates only room-specific candidate logits, progressively correcting cross-slot style conflicts as the global scene context evolves. Experiments on 3D-FRONT demonstrate state-of-the-art furniture retrieval and scene-level style coherence, producing more coherent fixed-layout furniture arrangements than object- and scene-level retrieval baselines.
Introduction
StyleForge formulates fixed-layout indoor furniture styling as scene-level structured selection, addressing conflicts that arise when independently suitable assets are composed. It uses a dynamic hypergraph style field, counterfactual preference learning, and iterative candidate refinement to improve scene-level coherence.
- Problem: Fixed-layout styling assigns one asset to each furniture slot while preserving category, position, orientation, and scale, making scene compatibility essential.An asset matching the target style in isolation may become incompatible when placed with other furniture.
- Motivation: Room style emerges from spatial, functional, and global relations, so candidate preferences must depend on choices at other furniture slots.Each local change can alter the global style field and reshape other slots’ preferences.
- Method: StyleForge maintains candidate distributions and uses style-conditioned hyperedge activation and propagation to capture local-to-global higher-order dependencies.A frozen multimodal large language model extracts structured room- and slot-level style priors from the style request and fixed layout.
- Method: Counterfactual style preference learning evaluates each candidate as a local substitution using candidate- and scene-level Mahalanobis energies for contextual compatibility.Learned dimension-wise tolerances distinguish restrictive style cues from weakly relevant variation.
- Optimization and Results: StyleForge alternates optimization of the style field and candidate logits, then performs test-time training that iteratively refines room-specific candidate distributions.Experiments on 3D-FRONT report state-of-the-art furniture retrieval and scene-level style coherence, supported by ablation and qualitative analyses.
Related Work
Prior work spans multimodal 3D asset retrieval and increasingly structured indoor scene synthesis. Recent research also develops higher-order hypergraph reasoning and energy-based structured prediction to model dependencies and compatibility beyond fixed pairwise relations.
- Indoor Scene Synthesis: Indoor scene synthesis has progressed from example-based arrangements and relation graphs to autoregressive generation and structured editing of layout, geometry, and appearance.Recent approaches use scene graphs, diffusion, and structured generation or editing on large furnished-scene datasets.
- Higher-Order Structured Inference: Scene graphs capture object relations, but fixed pairwise edges cannot fully represent groupwise material, color, form, and spatial relations underlying indoor style.Hypergraph neural networks and dynamic variants address multi-node relations by adapting higher-order connectivity to the input.
- Higher-Order Structured Inference: Recent structured inference work combines dynamic and multi-hop hypergraph reasoning, cross-node candidate generation, uncertainty modeling, and energy-based learning for compatible configurations.Related methods include multi-scale graph inference, distributional candidate generation, structured prediction energy networks, and iterative energy minimization.
Method
StyleForge jointly selects category-compatible assets for fixed furniture slots by maintaining candidate distributions and reasoning over room-level style coherence. Its dynamic hypergraph field and counterfactual Mahalanobis evaluation adapt to the target style, while alternating optimization refines scene-specific selections.
- Problem formulation: StyleForge assigns one category-compatible asset to each fixed slot while preserving prescribed category, position, orientation, and scale.The desired assignment must satisfy the style request both individually and as a complete scene.
- Structured priors and retrieval: A frozen MLLM extracts room- and slot-level style priors, and category-aware multimodal retrieval initializes a candidate distribution rather than a single asset.Retaining distributions lets scene-level reasoning revise locally plausible but globally incompatible choices.
- Dynamic hypergraph style field: Each slot node combines its soft asset representation, layout geometry, local and global style priors, and candidate uncertainty.The hypergraph uses spatially related local hyperedges and a complete-room global hyperedge whose participation depends on style and current candidate distributions.
- Dynamic hypergraph style field: Style-conditioned edge activation, strength, and attention allow the same fixed layout to induce different higher-order dependencies for different style requests.The propagated slot state captures contextual information while the global-edge state summarizes the room.
- Counterfactual preference learning: Counterfactual evaluation replaces one slot’s distribution with a candidate one-hot vector while keeping all other distributions fixed, then scores scene and candidate compatibility with Mahalanobis energies.Lower energy indicates better compatibility, and learned diagonal variances weight style dimensions.
- Optimization and inference: Training alternates style-field and candidate-logit updates; at inference, the MLLM and style field remain frozen while only room-specific logits are optimized for T steps.The final assignment selects the highest-probability candidate as distributions and global context co-evolve to correct initial retrieval conflicts.
Experiments
Experiments on 3D-FRONT evaluate StyleForge under fixed-layout styling constraints using slot-level retrieval accuracy and GPT-4o scene-level coherence assessments. Results show stronger scene-level retrieval, robust performance across room types, improved global compatibility through the proposed components, and progressive test-time refinement aligned with professional judgment.
- Dataset: Experiments use 7,100 training rooms and 800 test rooms across bedrooms, living rooms, dining rooms, and libraries/studies while preserving furniture categories, positions, orientations, and scales.Only the selected assets vary, creating fixed-layout indoor styling instances.
- Retrieval and scene-level evaluation: StyleForge improves over the strongest scene-aware baseline, MetaFind, by 34.6 R@1 points, while scene-aware retrieval substantially outperforms object-level retrieval.The comparison uses slot-level top-1 retrieval accuracy and GPT-4o ratings for aesthetic, color and material, style coherence, and realism and geometry.
- Room-type analysis: StyleForge remains effective across room types, with bedrooms comparatively easier and living rooms more challenging because of denser furniture relations.The results indicate that joint refinement continues correcting cross-slot compatibility as scene complexity increases.
- Qualitative comparison: Qualitative comparisons show that independent retrieval can produce mutually inconsistent colors, materials, silhouettes, scales, and spatial roles, whereas StyleForge favors assets supporting the global composition.Examples include visually dominant floor lamps, incompatible table–chair silhouettes, and mismatched cabinets and accessories.
- Ablation study: The ablation study attributes StyleForge’s performance to complementary higher-order context propagation, iterative candidate redistribution, and Mahalanobis-based style-dependent tolerance.The retrieval prior captures text–asset relevance, while the dynamic hypergraph, iterative updates, and Mahalanobis energy address composition-level conflicts and style variation.
- Test-time training and validation: During test-time training, reference-candidate probability, retrieval accuracy, and scene coherence increase together as multiple slots are progressively redistributed and asynchronously reordered.Professional approval also increases monotonically from Low to High GPT-4o AES groups under blinded evaluation by ten interior-design-experienced evaluators.
Conclusion
StyleForge performs scene-level fixed-layout furniture selection by modeling higher-order dependencies, evaluating contextual substitutions, and refining candidate distributions at test time. Its current optimization is limited by the initial retriever’s candidate recall, motivating reflection-guided iterative retrieval.
- Conclusion: StyleForge models higher-order furniture dependencies with a dynamic hypergraph style field, scores context-dependent substitutions using counterfactual Mahalanobis energies, and jointly refines candidate distributions at test time.It is presented as a scene-level structured selection framework for fixed-layout indoor furniture styling.
- Conclusion: StyleForge’s performance depends on the initial retriever’s recall because omitted assets cannot be reconsidered during subsequent scene-level optimization.The current method optimizes over a fixed candidate set.
- Conclusion: Future work will use reflection-guided iterative retrieval to update slot queries and candidate sets, reintroducing compatible assets missed during initialization.Scene-level feedback would dynamically update each slot’s retrieval query and candidate set.
StyleForge: Bridging Hypergraph Field and Counterfactual Reasoning for Indoor
The supplementary material documents StyleForge from furniture-category constraints and structured style-prior construction through scene-level evaluation. It covers category mappings and candidate constraints, followed by room- and slot-level style-prior extraction.
- The supplement provides a documentation chain spanning furniture-category constraints, structured style-prior construction, and scene-level evaluation.
- Appendix A: Appendix A maps coarse 3D-FRONT furniture categories to StyleForge’s fine-grained categories and clarifies candidate constraints under a fixed layout.
- Appendix B: Appendix B details extraction of structured room- and slot-level style priors with a frozen Qwen3-VL mode.
A. Furniture Category Mapping
StyleForge preserves each furniture slot’s dataset-provided coarse category and fixed layout while replacing only its asset. It refines coarse categories into validated fine-grained categories that directly determine each slot’s retrieval pool.
- Category and Layout Constraints: Dataset-provided coarse categories define fixed furniture slots, while retrieval and optimization preserve each slot’s category, position, orientation, and scale.StyleForge replaces only the asset assigned to each slot.
- Fine-Grained Taxonomy: StyleForge introduces a restricted fine-grained taxonomy to distinguish functionally and visually different furniture within coarse categories.For example, tables include coffee, dining, desk, and side tables, while chairs include dining, office, and lounge chairs.
- Category-Conditioned Retrieval: Qwen selects and validates a fine-grained category for each slot, and the matching taxonomy annotation directly determines its candidate retrieval pool.The predicted category is accepted only when it belongs to the restricted set associated with the slot’s dataset-provided coarse category.
B. Structured Style-Prior Extraction
StyleForge uses a frozen Qwen3-VL model to extract structured room- and slot-level style priors from the target style and fixed layout. Shared room descriptors guide category-specific asset queries while preserving fixed slot attributes and room-level coherence.
- Layout and constraints: A fixed-layout JSON specifies room walls and each furniture slot’s category, position, orientation, and size; geometry informs spatial function but is excluded from retrieval queries.The category, position, orientation, and size of every slot remain unchanged.
- Room- and slot-level priors: Qwen3-VL first generates shared room priors for palette, materials and finishes, and form language, then produces a category-specific query for every furniture slot.This shared context adapts room style to individual furniture categories.
- Query construction: Each retrieval query preserves the original fine-grained category and includes one style descriptor, one or two color descriptors, one material or finish, and one form descriptor.Queries must be complete positive English sentences rather than keyword lists, and must contain the input category verbatim.
- Cross-slot consistency: All slots share coherent room-level colors, materials, and form language, while slot-specific descriptions cannot alter their categories or layout attributes.The output preserves the fixed slot structure and avoids conflicting visual properties across slots.
- Conditioning the style field: The resulting room- and slot-level descriptions are embedded as z_r and z_i, respectively, and condition the dynamic hypergraph style field.For each slot, retrieval_text also serves as the candidate-asset query.
C. Scene-Level Evaluation Protocol
Scene-level quality is evaluated by rendering each method’s selected assets under identical conditions and having GPT-4o score four complementary dimensions. A blinded professional evaluation further assesses whether GPT-4o aesthetic ratings correspond to expert judgments.
- Scene construction: Each method’s selected assets are assembled with prescribed categories, positions, orientations, and scales, then rendered under identical camera, lighting, and rendering settings.GPT-4o receives the target-style description and corresponding rendered scene.
- GPT-4o evaluation: GPT-4o assigns integer scores from 1 to 5 for Aesthetic (AES), Color & Material (C&M), Style Coherence (SC), and Realism & Geometry (R&G).The dimensions cover visual quality, cross-furniture color and material coordination, target-style agreement and internal consistency, plus realism, scale, form, and geometry.
- GPT-4o evaluation: Five independent GPT-4o evaluations per scene are averaged arithmetically for AES, C&M, SC, and R&G, while textual reasons only verify reliance on intended visual factors.GPT-4o is used only for final evaluation and not for candidate retrieval, StyleForge training, or test-time adaptation.
- Professional validation: A blinded study evaluates 108 frozen scenes, stratified into Low, Medium, and High GPT-4o AES groups with 36 scenes each and judged by ten interior-design evaluators.Evaluators inspect randomized scenes with the target-style description and complete render; approval requires at least seven “Yes” judgments, with Wilson 95% confidence intervals reported.