# StyleForge：基于超图反事实推理的室内家具风格化

- 来源：HuggingFace Daily Papers（社区热门论文）
- 发布时间：2026-08-03 08:00
- AIHOT 分数：57
- AIHOT 链接：https://aihot.virxact.com/items/cmse2ervc0aabro2ecerhyqor
- 原文链接：https://arxiv.org/abs/2608.01954

## AI 摘要

StyleForge 提出一种场景级结构化选择框架，通过动态超图风格场为固定布局的室内家具挑选风格协调的资产。该方法利用冻结的多模态大语言模型提取风格先验，并采用反事实风格偏好学习，以马氏距离能量评估候选家具的上下文兼容性。在 3D-FRONT 上，StyleForge 在家具检索和场景级风格一致性上达到最优，优于物体级和场景级检索基线。

## 正文

Lingwei Dang

, Shishuo Shang

, Pan Liu

, Jiajia Cheng

, Ziyan Qiu

, Zhenhao Zhang

Yufei Zhu

, Shenghui Huang

, Qingxin Xiao

, Yun Hao

, Juntong Li

, Qingyao Wu

Abstract

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.

Figure 1: Comparison of two fixed-layout furniture-styling paradigms. (a) Independent object-level matching can combine individually relevant assets into a scene with shape, material, and color conflicts. (b) StyleForge jointly optimizes all slots to produce a coherent room-level assignment.

Introduction

An asset that matches a target style in isolation may become conspicuously incompatible when placed with other furniture in a three-dimensional room. This observation motivates fixed-layout indoor furniture styling: given a target-style description, a prescribed room layout, and a large asset library, the system must assign one asset to each furniture slot while preserving its category, position, orientation, and scale. Unlike general indoor scene synthesis, this task cannot avoid an incompatible combination by changing the spatial arrangement. Its fundamental challenge is therefore to bridge object-level semantic relevance and scene-level aesthetic coherence. Addressing this challenge is important for virtual interior design, 3D content creation, and immersive embodied environments.

Existing approaches mainly follow two paradigms. Object-level cross-modal retrieval methods use pretrained representations such as CLIP, ULIP, and OpenShape (Radford et al. 2021; Xue et al. 2023; Liu et al. 2023) to measure the similarity between a style description and an individual asset. Although these representations retrieve semantically relevant furniture, they cannot determine whether an asset remains appropriate after scene composition, where quality also depends on its color, material, form, and spatial relation to surrounding furniture. Relation-based methods instead model a scene with graphs or hypergraphs (Wald et al. 2020; Lv et al. 2024; Feng et al. 2019; Jiang et al. 2019). However, most rely on fixed pairwise edges or a static topology and therefore cannot express higher-order furniture dependencies that vary with the layout and requested style. Both paradigms also typically terminate after a single retrieval or ranking pass, without revising local choices according to the evolving scene context.

Our key insight is that an isolated furniture asset carries only partial style semantics, whereas room style emerges from the spatial, functional, and global relations among multiple assets. Fixed-layout furniture styling should therefore be formulated as scene-level structured selection rather than a collection of independent retrieval problems. Under this formulation, the preferred candidate for one slot depends on the current choices at other slots, while each local change alters the global style context.

Based on this insight, we propose StyleForge, a counterfactual reasoning framework built on a dynamic hypergraph style field. A frozen multimodal large language model extracts structured room- and slot-level style priors from the open-ended request and fixed layout. Instead of committing each slot to a top-ranked asset, StyleForge maintains a distribution over its candidates throughout optimization, allowing scene context to continuously reshape local preferences. The dynamic hypergraph style field represents furniture slots as distributional nodes and conditionally activates and weights layout-induced hyperedges according to the target style. Local hyperedges capture spatially or functionally related furniture groups, while a global hyperedge summarizes the room, enabling higher-order style dependencies to emerge from local-to-global propagation.

To convert scene compatibility into candidate preferences, we introduce counterfactual style preference learning. Each candidate is treated as a local substitution in the current style field while the remaining slot distributions are held fixed. Candidate- and scene-level Mahalanobis energies then measure the contextual compatibility of the resulting configuration. Their learned dimension-wise tolerances distinguish restrictive style cues from weakly relevant variation, providing an adaptive measure of conflicts in color, material, and form. StyleForge alternates between two optimization stages during training: with candidate distributions fixed, preference ranking learns an energy field that assigns lower energy to coherent configurations; with the field fixed, the resulting counterfactual energies update the candidate logits. At inference, the multimodal large language model and learned style field remain frozen, and test-time training updates only the room-specific candidate logits. The furniture assignment can consequently improve with its evolving global context instead of terminating after one-shot retrieval.

Experiments on 3D-FRONT show that StyleForge achieves state-of-the-art furniture retrieval and scene-level style coherence. Ablation and qualitative analyses further support the effectiveness of dynamic higher-order reasoning, Mahalanobis energy, and iterative refinement.

1.

We introduce a dynamic hypergraph style field that represents furniture slots as candidate distributions and uses style-conditioned hyperedge activation and propagation to capture local-to-global higher-order style dependencies.

2.

We propose counterfactual style preference learning, which evaluates each candidate as a local substitution in the current style field and uses candidate- and scene-level Mahalanobis energies to measure its contextual compatibility.

3.

We develop an alternating optimization and test-time training strategy that learns the style field and candidate logits in alternating stages, then freezes the field and iteratively refines room-specific candidate distributions at inference.

Related Work

3D Asset Retrieval.

Vision–language pretraining provides shared semantic spaces for open-vocabulary 3D asset retrieval. CLIP establishes transferable image–text representations, while ULIP, ULIP-2, OpenShape, and Uni3D align language, images, and point clouds for scalable 3D representation learning (Radford et al. 2021; Xue et al. 2023, 2024; Liu et al. 2023; Zhou et al. 2024); Point-Bind and OpenScene extend multimodal alignment to richer 3D understanding (Guo et al. 2023; Peng et al. 2023). These methods primarily score text–asset relevance and cannot determine whether locally plausible furniture remains coherent after composition in color, material, and form. Prior work learns style from furniture compatibility, real-world co-occurrence, or scene images (Liu et al. 2015, 2019; Weiss et al. 2020), and exploits local context or joint layout–object encodings for recommendation (Savva et al. 2017; Pan et al. 2025). StyleForge instead represents all slots as coupled candidate distributions and jointly optimizes the complete assignment under a shared higher-order style field.

Figure 2: Overview of StyleForge. A frozen MLLM generates structured style priors and initializes candidate distributions. A dynamic hypergraph propagates higher-order context, while counterfactual substitutions are scored by Mahalanobis energy. Training alternates field and logit updates; TTT optimizes room-specific candidate logits.

Indoor Scene Synthesis.

Indoor scene synthesis has evolved from example-based arrangements and relation-graph modeling (Fisher et al. 2012; Wang et al. 2019) to autoregressive generation on large furnished-scene datasets (Fu et al. 2021; Paschalidou et al. 2021). Recent approaches use scene graphs, diffusion, and structured generation or editing to model layout, geometry, and appearance (Zhai et al. 2023; Tang et al. 2024; Ju et al. 2024; Zhao et al. 2024; Lin and MU 2024; Zhai et al. 2024). Large language and vision–language models further support instruction following, commonsense constraints, and style-conditioned planning (Yang et al. 2024; Feng et al. 2023; Sun et al. 2025; Marshall et al. 2025; Pan and Liu 2026; Berdoz et al. 2026), while related work studies multimodal generation, implicit representations, editable synthesis, and texturing (Yang et al. 2025; Liang et al. 2025; Zheng et al. 2025; Höllein et al. 2023; Huang et al. 2025). These methods generate or modify object categories, poses, geometry, or appearance. We address a complementary, more constrained problem: selecting real library assets that jointly realize a target style while categories, positions, orientations, and scales remain fixed.

Higher-Order Structured Inference.

Scene graphs encode object relations for 3D scene understanding (Wald et al. 2020; Lv et al. 2024), but fixed pairwise edges cannot fully express the groupwise material, color, form, and spatial relations that define indoor style. Hypergraph neural networks model multi-node relations, and dynamic variants adapt higher-order connectivity to the input (Feng et al. 2019; Jiang et al. 2019). Recent work further studies dynamic and multi-hop hypergraph reasoning (Zhou et al. 2023; Xie et al. 2025; Li et al. 2025); in other structured prediction domains, multi-scale graph inference and distributional candidate generation model cross-node dependencies and output uncertainty (Dang et al. 2021, 2022). In parallel, energy-based learning assigns low energy to compatible configurations (LeCun et al. 2006). Structured prediction energy networks and iterative energy minimization support global inference over coupled outputs (Belanger and McCallum 2016; Du et al. 2022), while compositional energy models combine multiple constraints (Du et al. 2020). StyleForge applies these ideas to style-aware asset selection: the target style activates and weights layout-induced hyperedges, nodes represent candidate distributions rather than fixed features, and learned Mahalanobis energies evaluate counterfactual local substitutions.

Method

Problem Formulation and Overview

Given a fixed indoor layout, a target-style description, and an asset library , our goal is to assign one asset to each of furniture slots while preserving the prescribed category, position, orientation, and scale of every slot. Slot is specified by its category and geometric descriptor . After category filtering, we retrieve a candidate set . The desired output is a joint assignment that agrees with the style request both individually and as a scene.

StyleForge instead maintains a categorical candidate distribution for every slot and optimizes all distributions jointly. As shown in Fig. 2, a frozen MLLM extracts structured room- and slot-level style priors from the request and layout. A dynamic hypergraph style field then propagates higher-order context, while counterfactual Mahalanobis energies evaluate each candidate as a local substitution. Training alternates between learning the style field and refining candidate logits; at test time, all model parameters are frozen and only the room-specific candidate logits are updated.

Style-Prior Candidate Initialization

Structured style prior. We prompt the frozen MLLM with the target description and fixed layout to produce room-level and slot-specific retrieval descriptions. Their embeddings and encode the shared palette, material, and form language and its category-specific realization, respectively.

Category-aware retrieval. A frozen multimodal encoder indexes multi-view asset renders. For slot , category-constrained retrieval returns the top- candidates and similarities , from which we initialize

(1)

where is the retrieval temperature, controls initialization sharpness, and ensures numerical stability. Retaining a distribution rather than a single retrieval allows scene-level reasoning to revise locally plausible but globally incompatible choices.

Dynamic Hypergraph Style Field

Distributional slot nodes. Each node represents a furniture slot . For the candidate asset denotes its fixed visual-semantic feature. Given , its soft asset representation, uncertainty, and initial state are

(2)

where is a learnable node encoder. Thus, each state combines the current selection, layout, local and global style priors, and candidate uncertainty.

Style-conditioned propagation. Following hypergraph message passing, we construct from the fixed layout. Local hyperedges connect spatially related slots, and a global hyperedge contains the complete room. Each edge has a structural descriptor . The layout fixes the candidate topology, whereas the target style and current candidate distributions determine which edges participate at each layer.

Let be the state of edge at layer . We predict its activation probability , aggregate incident nodes with normalized attention , and update the edge state:

(3)

Let be the effective edge weight. Using attention normalized over edges incident to node , edge-to-node propagation becomes

(4)

After layers, captures the contextual state of slot , while the global-edge state summarizes the room. Because edge activation, strength, and attention depend on the style and current distributions, the same layout can induce different higher-order dependencies for different requests.

Counterfactual Style Preference Learning

A candidate that matches the target text in isolation may still conflict with the scene. We therefore evaluate candidates through counterfactual substitution. For candidate , the distribution of slot is replaced by its one-hot vector, , while all other slot distributions , , remain fixed. This evaluates alternatives in a shared scene context, making their compatibility directly comparable.

We evaluate scene- and candidate-level compatibility.For the scene-level energy, we project the room state as and predict a diagonal Gaussian prototype from . For the candidate-level energies, we contextualize each candidate as and predict from . We define the mean diagonal Mahalanobis energy as

(5)

The scene- and candidate-level energies are instantiated as and , respectively. Thus, scores the joint configuration, whereas scores candidate at slot . The learned diagonal variances weight style dimensions, and lower energy indicates better compatibility.

To avoid degenerate minima, we rank ground-truth configurations below random or embedding-similar substitutions:

(6)

where is the ranking margin, controls the strength of the sparsity regularization, and discourages indiscriminate edge activation.

Training and Inference Strategy

During training, StyleForge alternates between optimizing the style-field parameters and the training-room candidate logits . With the candidate distributions fixed, we update using the preference-ranking objective in Eq. (6); with fixed, we update the candidate logits by minimizing

(7)

媒体内容 · 前往原文查看

Method Init. R@1 Final R@1 AES C&M SC R&G Avg.

GT – – 4.58 4.64 4.62 4.67 4.63

ULIP [CVPR 2023] – 33.9 3.02 3.08 2.86 3.18 3.04

OpenShape [NeurIPS 2023] – 35.1 3.10 3.14 3.02 3.06 3.08

Uni3D [ICLR 2024] – 36.3 3.17 3.23 3.21 2.95 3.14

MetaFind [NeurIPS 2025] – 44.5 4.19 4.27 4.18 4.34 4.25

StyleForge (Ours) 22.8 79.1 4.53 4.61 4.58 4.64 4.59

Table 1: Comparison with 3D asset retrieval methods on 3D-FRONT. AES, C&M, SC, and R&G denote Aesthetic, Color & Material, Style Coherence, and Realism & Geometry, respectively. Higher is better for all reported metrics. Light blue denotes our method, and bold indicates the best comparable result in each metric column.

Each training round alternates style-field updates with logit updates, exposing the style field to the candidate-distribution shifts induced by iterative optimization. Algorithm 1 summarizes this alternating procedure. During inference, we initialize using Eq. (1), freeze the MLLM and , and optimize only the room-specific candidate logits for steps. The final assignment selects . As the distributions and global scene context co-evolve, StyleForge progressively corrects conflicts introduced by the initial retrieval.

媒体内容 · 前往原文查看

Algorithm 1 Alternating Training of StyleForge

Input: Training rooms , asset library , and frozen MLLM Output: Style-field parameters and training-room logits

1: Generate priors and candidates; initialize using Eq. (1)

2: for each alternating training round do

3: Freeze

4: for to do

5: Construct positive and random/hard-negative configurations

6: Update using Eq. (6)

7: end for

8: Freeze

9: for each room do

10: Update for steps using Eq. (7)

11: end for

12: end for

13: return and

Experiments

Experimental Setup

Dataset. We conduct experiments on 3D-FRONT (Fu et al. 2021), using 7,100 rooms for training and 800 rooms for testing. The test set covers bedrooms, living rooms, dining rooms, and libraries/studies. For each room, we preserve the original furniture categories, positions, orientations, and scales and vary only the selected assets, thereby constructing fixed-layout indoor styling instances.

Evaluation metrics. We report slot-level top-1 retrieval accuracy. A prediction is considered correct when the selected asset matches the reference asset for that slot. Init. R@1 denotes the accuracy of StyleForge’s initial retrieval prior, whereas Final R@1 denotes the accuracy of the final prediction produced by each method. We additionally use GPT-4o to assess four scene-level properties on a five-point scale. Aesthetic (AES) evaluates overall visual quality, balance, and atmosphere. Color & Material (C&M) measures the coordination of colors, textures, and materials across furniture. Style Coherence (SC) evaluates agreement with the target-style description and consistency among the selected assets. Realism & Geometry (R&G) assesses visual realism, scale compatibility, and geometric plausibility. GPT-4o receives the target-style description and the rendered scene and evaluates each scene independently five times. The five scores are then averaged. GPT-4o is used only for evaluation and does not participate in candidate retrieval, training, or TTT.

Baselines. We compare StyleForge with representative 3D asset retrieval methods. ULIP (Xue et al. 2023), OpenShape (Liu et al. 2023), and Uni3D (Zhou et al. 2024) measure object-level text–asset relevance. MetaFind (Pan et al. 2025) additionally encodes existing objects and layout context, providing a stronger scene-aware baseline. Because MetaFind has not released its source code, we reproduce its architecture, training objective, and retrieval procedure following the paper. All methods use the same data split, asset library, category constraints, fixed layouts, target-style descriptions, and top-15 candidate sets.

Implementation. We instantiate the frozen MLLM with Qwen3-VL (Bai et al. 2025) to generate room- and slot-level style priors and use the frozen Qwen3-VL-Embedding model (Li et al. 2026) to retrieve fifteen category-compatible candidates per slot. Neither foundation model is updated during training or inference. Training is conducted on a single NVIDIA A800-SXM4-80GB GPU. The dynamic hypergraph and energy field are trained using the alternating procedure. During each alternating cycle, we perform one model-parameter update followed by five candidate-logit updates, using learning rates of and , respectively. At inference, the MLLM, multimodal encoder, and learned style field remain frozen; only room-specific candidate logits are updated for 200 TTT steps.

Figure 3: Qualitative comparison under identical layouts and target styles. Red boxes indicate selected assets that are inconsistent with the overall scene style. StyleForge jointly selects assets to reduce such style conflicts while preserving the prescribed layout.

Quantitative Results

Table 1 shows that scene-aware retrieval substantially outperforms object-level retrieval, while StyleForge further improves over the strongest scene-aware baseline, MetaFind, by 34.6 R@1 points. Object-level methods judge whether each asset matches the text prompt in isolation and can therefore select furniture pieces that are individually plausible but mutually inconsistent once composed in a fixed layout, where conflicts in color, material, form, scale, and visual balance become apparent. MetaFind alleviates this limitation by incorporating existing objects and layout context, explaining its advantage over ULIP, OpenShape, and Uni3D. However, its ranking remains conditioned on local or partial-scene context and cannot jointly revise all slot decisions. StyleForge instead formulates fixed-layout styling as a coupled structured selection problem: each slot distribution shapes the global style field, which in turn re-evaluates the contextual compatibility of every local candidate. This bidirectional interaction suppresses assets that match the prompt in isolation but disrupt the composed room.

媒体内容 · 前往原文查看

Room Type Avg. S. Init. R@1 Final R@1 AES

Bedroom 5.2 24.6 82.4 4.58

Living Room 8.4 20.9 75.8 4.47

Dining Room 6.1 23.4 80.6 4.55

Library/Study 4.7 22.1 78.3 4.50

Overall 6.3 22.8 79.1 4.53

Table 2: StyleForge performance by room type. Avg. S. is the average number of furniture slots per room.

Table 2 further shows that StyleForge performs consistently across room types. Bedrooms generally contain fewer slots and thus induce a smaller joint assignment space, which makes them comparatively easier. Living rooms are more challenging because they contain more furniture and denser functional and visual relations, yet StyleForge still produces a substantial improvement. The stable behavior across these settings indicates that joint refinement continues to propagate and correct cross-slot compatibility as scene complexity increases. It also suggests that the dynamic hypergraph style field captures higher-order dependencies induced jointly by layout, functional relations, and the target style rather than memorizing a particular room template.

Qualitative Analysis

Figure 3 compares all methods under identical layouts and target styles. Object-level baselines often retrieve assets that match the prompt individually but become visually disruptive in the composed scene, such as a floor lamp with an excessively high visual center, table–chair combinations with incompatible silhouette languages, or cabinets and accessories whose materials and colors do not correspond. These failures show that fixed-layout styling cannot rely solely on independent text–asset relevance because coherence emerges from relative scale, repeated forms, material correspondence, and spatial roles across furniture. StyleForge instead favors assets that support one another in the global composition, aligning primary furniture, secondary pieces, and decorative elements through a shared formal language and visual rhythm.

Ablation Study

媒体内容 · 前往原文查看

HG Iter. Maha. Final R@1 AES

22.8 2.88

✓ ✓ 23.7 2.91

✓ ✓ 61.7 4.37

✓ ✓ ✓ 79.1 4.53

Table 3: Ablation study. HG, Iter., Maha., and AES denote the dynamic hypergraph, iterative update, Mahalanobis energy, and Aesthetic score, respectively. The first row uses only the retrieval prior; without Maha., Euclidean energy is used.

Table 3 verifies the complementary roles of the main components. The retrieval prior captures text–asset relevance but cannot determine whether individually plausible candidates conflict after composition. Adding the dynamic hypergraph and Mahalanobis energy provides a scene-level compatibility measure, yet without iterative refinement the model cannot propagate a local correction through the coupled slot distributions. Retaining the hypergraph and iterative updates while replacing Mahalanobis energy with Euclidean energy enables joint reassignment, but the isotropic metric treats all style dimensions as equally important and cannot distinguish style-defining attributes from acceptable variation. The full model combines higher-order context propagation, progressive candidate redistribution, and style-dependent tolerance, allowing it to identify which deviations disrupt scene coherence and which remain compatible with the target style.

Convergence Analysis

媒体内容 · 前往原文查看

Steps Final R@1 GT Prob. SC Changed Slots

0 22.8 11.9 2.92 0.0

20 38.6 23.7 3.34 1.4

50 55.9 42.1 3.82 2.7

100 70.8 67.2 4.25 3.6

200 79.1 84.6 4.58 4.1

Table 4: Round-prefix convergence of StyleForge TTT. GT Prob. is the average probability assigned to the ground-truth candidate; Changed Slots is the average number of slots whose top-ranked candidate differs from the initial retrieval.

Figure 4: Iterative TTT on a five-slot room. Colored boxes mark initial or first-time GT selections.

Table 4 and Figure 4 illustrate the dynamics of test-time training. As optimization proceeds, the probability assigned to reference candidates, retrieval accuracy, and scene coherence increase together, while the number of slots that change their top-ranked candidate also grows. The gains therefore arise from progressive joint redistribution across multiple slots rather than a one-shot replacement at a single location. The plotted ground-truth probabilities are used only as post-hoc diagnostics and are never observed by the optimizer.

A key observation is that different slots converge asynchronously. Once the selection at one slot changes, the global style field is updated; related candidates at other slots then receive new compatibility estimates and may be reordered in subsequent steps. This chain of corrections exposes the central difficulty of fixed-layout furniture styling: locally optimal assets do not necessarily form a globally coherent composition, and the benefit of a local substitution may become visible only under the updated scene context. By continuing to optimize candidate logits at inference time, StyleForge turns this context dependence into an explicit iterative selection process, consistent with counterfactual preference learning that evaluates each candidate as a local substitution within the current room.

Professional Validation

To examine whether the predicted AES aligns with professional judgment, we divide scenes into Low, Medium, and High AES groups and ask ten evaluators with interior-design experience to assess them under a blinded protocol. Table 5 shows that professional approval increases monotonically from the Low group to the High group, with a clear separation between the two extremes. This association suggests that GPT-4o AES is not an isolated numerical indicator but reflects perceptual properties emphasized by professional evaluators, including overall visual quality, coordination, and scene completeness. Scene-level automatic evaluation can therefore serve as a useful complement to expert judgment when assessing fixed-layout furniture styling.

媒体内容 · 前往原文查看

AES Level Approved Rate 95% CI

Low 8/36 22.2% [11.7%, 38.1%]

Medium 21/36 58.3% [42.2%, 72.9%]

High 31/36 86.1% [71.3%, 93.9%]

Table 5: Professional approval by predicted AES level. Approval requires positive judgments from at least seven of the ten blinded evaluators; confidence intervals (CIs) are Wilson 95% intervals.

Conclusion

We present StyleForge, a scene-level structured selection framework for fixed-layout indoor furniture styling. Rather than treating furniture as independent retrieval targets, it models higher-order dependencies with a dynamic hypergraph style field, scores context-dependent substitutions through counterfactual Mahalanobis energies, and jointly refines candidate distributions at test time. The results highlight a central insight: style coherence arises from coordination in color, material, shape, and spatial role, so locally optimal assets need not form a globally coherent scene and must be revised against evolving context. Experiments on 3D-FRONT confirm the complementary benefits of higher-order modeling, contextual energy evaluation, and iterative inference for asset recovery and scene quality. StyleForge currently optimizes over a fixed candidate set, making its performance dependent on the recall of the initial retriever: assets omitted from the initial top- set cannot be reconsidered during subsequent scene-level optimization. Future work will explore reflection-guided iterative retrieval, using scene-level feedback to dynamically update each slot’s retrieval query and candidate set so that compatible assets missed during initialization can be reintroduced into the optimization.

References

S. Bai, Y. Cai, R. Chen, K. Chen, X. Chen, Z. Cheng, L. Deng, W. Ding, C. Gao, C. Ge, W. Ge, Z. Guo, Q. Huang, J. Huang, F. Huang, B. Hui, S. Jiang, Z. Li, M. Li, M. Li, K. Li, Z. Lin, J. Lin, X. Liu, J. Liu, C. Liu, Y. Liu, D. Liu, S. Liu, D. Lu, R. Luo, C. Lv, R. Men, L. Meng, X. Ren, X. Ren, S. Song, Y. Sun, J. Tang, J. Tu, J. Wan, P. Wang, P. Wang, Q. Wang, Y. Wang, T. Xie, Y. Xu, H. Xu, J. Xu, Z. Yang, M. Yang, J. Yang, A. Yang, B. Yu, F. Zhang, H. Zhang, X. Zhang, B. Zheng, H. Zhong, J. Zhou, F. Zhou, J. Zhou, Y. Zhu, and K. Zhu (2025) Qwen3-vl technical report. External Links: 2511.21631, Link Cited by: Experimental Setup.

D. Belanger and A. McCallum (2016) Structured prediction energy networks. In ICML, pp. 983–992. External Links: Link Cited by: Higher-Order Structured Inference..

F. Berdoz, L. A. Lanzendörfer, N. Tuninga, and R. Wattenhofer (2026) Text-to-scene with large reasoning models. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 40, pp. 2435–2443. Cited by: Indoor Scene Synthesis..

L. Dang, Y. Nie, C. Long, Q. Zhang, and G. Li (2021) Msr-gcn: multi-scale residual graph convolution networks for human motion prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pp. 11467–11476. Cited by: Higher-Order Structured Inference..

L. Dang, Y. Nie, C. Long, Q. Zhang, and G. Li (2022) Diverse human motion prediction via gumbel-softmax sampling from an auxiliary space. In Proceedings of the 30th ACM international conference on multimedia, pp. 5162–5171. Cited by: Higher-Order Structured Inference..

Y. Du, S. Li, and I. Mordatch (2020) Compositional visual generation and inference with energy based models. External Links: 2004.06030, Link Cited by: Higher-Order Structured Inference..

Y. Du, S. Li, J. B. Tenenbaum, and I. Mordatch (2022) Learning iterative reasoning through energy minimization. In ICML, pp. 5570–5582. External Links: Link Cited by: Higher-Order Structured Inference..

W. Feng, W. Zhu, T. Fu, V. Jampani, A. Akula, X. He, S. Basu, X. E. Wang, and W. Y. Wang (2023) Layoutgpt: compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems 36, pp. 18225–18250. Cited by: Indoor Scene Synthesis..

Y. Feng, H. You, Z. Zhang, R. Ji, and Y. Gao (2019) Hypergraph neural networks. In Proceedings of the AAAI conference on artificial intelligence, Vol. 33, pp. 3558–3565. Cited by: Introduction, Higher-Order Structured Inference..

M. Fisher, D. Ritchie, M. Savva, T. Funkhouser, and P. Hanrahan (2012) Example-based synthesis of 3d object arrangements. ACM Transactions on Graphics (TOG) 31 (6), pp. 1–11. Cited by: Indoor Scene Synthesis..

H. Fu, B. Cai, L. Gao, L. Zhang, J. Wang, C. Li, Q. Zeng, C. Sun, R. Jia, B. Zhao, et al. (2021) 3d-front: 3d furnished rooms with layouts and semantics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 10933–10942. Cited by: Indoor Scene Synthesis., Experimental Setup.

Z. Guo, R. Zhang, X. Zhu, Y. Tang, X. Ma, J. Han, K. Chen, P. Gao, X. Li, H. Li, and P. Heng (2023) Point-bind & point-llm: aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. External Links: 2309.00615, Link Cited by: 3D Asset Retrieval..

L. Höllein, A. Cao, A. Owens, J. Johnson, and M. Nießner (2023) Text2room: extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 7909–7920. Cited by: Indoor Scene Synthesis..

Z. Huang, W. Yu, X. Cheng, C. Zhao, Y. Ge, M. Guo, L. Yuan, and Y. Tian (2025) Roompainter: view-integrated diffusion for consistent indoor scene texturing. In Proceedings of the Computer Vision and Pattern Recognition Conference, pp. 574–584. Cited by: Indoor Scene Synthesis..

J. Jiang, Y. Wei, Y. Feng, J. Cao, and Y. Gao (2019) Dynamic hypergraph neural networks.. In Ijcai, pp. 2635–2641. Cited by: Introduction, Higher-Order Structured Inference..

X. Ju, Z. Huang, Y. Li, G. Zhang, Y. Qiao, and H. Li (2024) Diffindscene: diffusion-based high-quality 3d indoor scene generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 4526–4535. Cited by: Indoor Scene Synthesis..

Y. LeCun, S. Chopra, R. Hadsell, M. Ranzato, F. Huang, et al. (2006) A tutorial on energy-based learning. Predicting structured data 1 (0). Cited by: Higher-Order Structured Inference..

C. Li, T. Li, X. Hu, D. Luo, and T. Jin (2025) Dvhgnn: multi-scale dilated vision hgnn for efficient vision recognition. In Proceedings of the Computer Vision and Pattern Recognition Conference, pp. 20158–20168. Cited by: Higher-Order Structured Inference..

M. Li, Y. Zhang, D. Long, K. Chen, S. Song, S. Bai, Z. Yang, P. Xie, A. Yang, D. Liu, J. Zhou, and J. Lin (2026) Qwen3-vl-embedding and qwen3-vl-reranker: a unified framework for state-of-the-art multimodal retrieval and ranking. External Links: 2601.04720, Link Cited by: Experimental Setup.

Z. Liang, G. Xu, H. Wu, Y. Huang, W. Li, and L. Duan (2025) S-inf: towards realistic indoor scene synthesis via scene implicit neural field. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39, pp. 5173–5181. Cited by: Indoor Scene Synthesis..

C. Lin and Y. MU (2024) InstructScene: instruction-driven 3d indoor scene synthesis with semantic graph prior. In The Twelfth International Conference on Learning Representations, External Links: Link Cited by: Indoor Scene Synthesis..

M. Liu, R. Shi, K. Kuang, Y. Zhu, X. Li, S. Han, H. Cai, F. Porikli, and H. Su (2023) Openshape: scaling up 3d shape representation towards open-world understanding. Advances in neural information processing systems 36, pp. 44860–44879. Cited by: Introduction, 3D Asset Retrieval., Experimental Setup.

T. Liu, A. Hertzmann, W. Li, and T. Funkhouser (2015) Style compatibility for 3d furniture models. ACM Transactions on Graphics (TOG) 34 (4), pp. 1–9. Cited by: 3D Asset Retrieval..

Y. Liu, R. Tang, and D. Ritchie (2019) Learning style compatibility between objects in a real-world 3d asset database. In Computer Graphics Forum, Vol. 38, pp. 775–784. Cited by: 3D Asset Retrieval..

C. Lv, M. Qi, X. Li, Z. Yang, and H. Ma (2024) Sgformer: semantic graph transformer for point cloud-based 3d scene graph generation. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38, pp. 4035–4043. Cited by: Introduction, Higher-Order Structured Inference..

K. O. Marshall, O. Poursaeed, S. Oprea, A. Kumar, A. Jignasu, C. Hegde, Y. Li, and R. Ranjan (2025) Decorum: a language-based approach for style-conditioned synthesis of indoor 3d scenes. arXiv preprint arXiv:2503.18155. Cited by: Indoor Scene Synthesis..

Z. Pan and H. Liu (2026) MetaSpatial: reinforcing 3d spatial reasoning in VLMs for the metaverse. In The Fourteenth International Conference on Learning Representations, External Links: Link Cited by: Indoor Scene Synthesis..

Z. Pan, Y. Lu, and H. Liu (2025) MetaFind: scene-aware 3d asset retrieval for coherent metaverse scene generation. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, External Links: Link Cited by: 3D Asset Retrieval., Experimental Setup.

D. Paschalidou, A. Kar, M. Shugrina, K. Kreis, A. Geiger, and S. Fidler (2021) ATISS: autoregressive transformers for indoor scene synthesis. In Proceedings of the 35th International Conference on Neural Information Processing Systems, NIPS ’21, Red Hook, NY, USA. External Links: ISBN 9781713845393 Cited by: Indoor Scene Synthesis..

S. Peng, K. Genova, C. Jiang, A. Tagliasacchi, M. Pollefeys, T. Funkhouser, et al. (2023) Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 815–824. Cited by: 3D Asset Retrieval..

A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. (2021) Learning transferable visual models from natural language supervision. In International conference on machine learning, pp. 8748–8763. Cited by: Introduction, 3D Asset Retrieval..

M. Savva, A. X. Chang, and M. Agrawala (2017) Scenesuggest: context-driven 3d scene design. arXiv preprint arXiv:1703.00061. Cited by: 3D Asset Retrieval..

F. Sun, W. Liu, S. Gu, D. Lim, G. Bhat, F. Tombari, M. Li, N. Haber, and J. Wu (2025) Layoutvlm: differentiable optimization of 3d layout via vision-language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pp. 29469–29478. Cited by: Indoor Scene Synthesis..

J. Tang, Y. Nie, L. Markhasin, A. Dai, J. Thies, and M. Nießner (2024) Diffuscene: denoising diffusion models for generative indoor scene synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 20507–20518. Cited by: Indoor Scene Synthesis..

J. Wald, H. Dhamo, N. Navab, and F. Tombari (2020) Learning 3d semantic scene graphs from 3d indoor reconstructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 3961–3970. Cited by: Introduction, Higher-Order Structured Inference..

K. Wang, Y. Lin, B. Weissmann, M. Savva, A. X. Chang, and D. Ritchie (2019) PlanIT: planning and instantiating indoor scenes with relation graph and spatial prior networks. ACM Trans. Graph. 38 (4). External Links: ISSN 0730-0301, Link, Document Cited by: Indoor Scene Synthesis..

T. Weiss, I. Yildiz, N. Agarwal, E. Ataer-Cansizoglu, and J. Choi (2020) Image-driven furniture style for interactive 3d scene modeling. In Computer Graphics Forum, Vol. 39, pp. 57–68. Cited by: 3D Asset Retrieval..

L. Xie, S. Gao, J. Liu, M. Yin, and T. Jin (2025) K-hop hypergraph neural network: a comprehensive aggregation approach. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39, pp. 21679–21687. Cited by: Higher-Order Structured Inference..

L. Xue, M. Gao, C. Xing, R. Martín-Martín, J. Wu, C. Xiong, R. Xu, J. C. Niebles, and S. Savarese (2023) Ulip: learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 1179–1189. Cited by: Introduction, 3D Asset Retrieval., Experimental Setup.

L. Xue, N. Yu, S. Zhang, A. Panagopoulou, J. Li, R. Martín-Martín, J. Wu, C. Xiong, R. Xu, J. C. Niebles, et al. (2024) Ulip-2: towards scalable multimodal pre-training for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 27091–27101. Cited by: 3D Asset Retrieval..

Y. Yang, F. Sun, L. Weihs, E. VanderBilt, A. Herrasti, W. Han, J. Wu, N. Haber, R. Krishna, L. Liu, et al. (2024) Holodeck: language guided generation of 3d embodied ai environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 16227–16237. Cited by: Indoor Scene Synthesis..

Z. Yang, K. Lu, C. Zhang, J. Qi, H. Jiang, R. Ma, S. Yin, Y. Xu, M. Xing, Z. Xiao, et al. (2025) Mmgdreamer: mixed-modality graph for geometry-controllable 3d indoor scene generation. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39, pp. 9391–9399. Cited by: Indoor Scene Synthesis..

G. Zhai, E. P. Örnek, D. Z. Chen, R. Liao, Y. Di, N. Navab, F. Tombari, and B. Busam (2024) Echoscene: indoor scene generation via information echo over scene graph diffusion. In European Conference on Computer Vision, pp. 167–184. Cited by: Indoor Scene Synthesis..

G. Zhai, E. P. Örnek, S. Wu, Y. Di, F. Tombari, N. Navab, and B. Busam (2023) Commonscenes: generating commonsense 3d indoor scenes with scene graph diffusion. Advances in Neural Information Processing Systems 36, pp. 30026–30038. Cited by: Indoor Scene Synthesis..

Y. Zhao, Z. Zhao, J. Li, S. Dong, and S. Gao (2024) RoomDesigner: encoding anchor-latents for style-consistent and shape-compatible indoor scene generation. In 2024 International Conference on 3D Vision (3DV), pp. 1413–1423. Cited by: Indoor Scene Synthesis..

K. Zheng, X. Chen, X. He, J. Gu, L. Li, Z. Yang, K. Lin, J. Wang, L. Wang, and X. E. Wang (2025) EditRoom: LLM-parameterized graph diffusion for composable 3d room layout editing. In The Thirteenth International Conference on Learning Representations, External Links: Link Cited by: Indoor Scene Synthesis..

J. Zhou, J. Wang, B. Ma, Y. Liu, T. Huang, and X. Wang (2024) Uni3D: exploring unified 3d representation at scale. In International Conference on Learning Representations, B. Kim, Y. Yue, S. Chaudhuri, K. Fragkiadaki, M. Khan, and Y. Sun (Eds.), Vol. 2024, pp. 46766–46782. External Links: Link Cited by: 3D Asset Retrieval., Experimental Setup.

P. Zhou, Z. Wu, X. Zeng, G. Wen, J. Ma, and X. Zhu (2023) Totally dynamic hypergraph neural network. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pp. 2476–2483. Cited by: Higher-Order Structured Inference..

StyleForge: Bridging Hypergraph Field and Counterfactual Reasoning for Indoor Furniture Styling Supplementary Material

Supplementary organization. This supplementary material provides a complete documentation chain from furniture-category constraints and structured style-prior construction to scene-level evaluation. Appendix A defines the mapping from the coarse furniture categories in 3D-FRONT to the fine-grained categories used by StyleForge and clarifies the candidate constraints under a fixed layout. Appendix B details the extraction of structured room- and slot-level style priors with a frozen Qwen3-VL model, including the complete prompting protocol. Appendix C specifies the unified scene-rendering setup, the GPT-4o-based four-dimensional scene-quality evaluation protocol, and the blinded professional validation procedure. For more vivid demonstrations of behavior that is difficult to convey with static figures, please refer to the supplementary videos provided alongside this document. The videos visualize the progressive evolution of candidate probabilities and furniture assignments during StyleForge test-time training and provide complete-scene comparisons with representative object-level and scene-aware retrieval methods.

A. Furniture Category Mapping

The 3D-FRONT dataset assigns each furniture instance a coarse category, such as sofa, bed, table, chair, cabinet/shelf/desk, lighting, stool, or other. In fixed-layout furniture styling, these dataset-provided categories define the furniture slots and remain unchanged throughout candidate retrieval and scene-level optimization. StyleForge replaces only the asset assigned to each slot while preserving its original category, position, orientation, and scale.

Because a coarse category can contain furniture with substantially different functions and appearances, we introduce a restricted fine-grained taxonomy for StyleForge. For example, the dataset-level table category includes coffee tables, dining tables, desks, and side tables, whereas the chair category includes dining chairs, office chairs, and lounge chairs. Table 6 lists the complete mapping from the dataset-provided coarse categories to the fine-grained categories used by StyleForge.

媒体内容 · 前往原文查看

Dataset Coarse Category StyleForge Fine-Grained Categories

Sofa Sofa; sectional sofa; two-seat sofa

Bed Bed; bunk bed

Table Table; coffee table; dining table; desk; side table; console table; bar table

Chair Chair; lounge chair; dining chair; office chair; bar chair; bench; armchair

Cabinet/Shelf/Desk Cabinet; TV stand; chest of drawers; sideboard; bookshelf; wardrobe; shoe cabinet; filing cabinet; vanity table; wine cabinet; display cabinet; nightstand

Lighting Floor lamp; table lamp; pendant lamp; ceiling lamp

Pier/Stool Stool; footstool; pouf

Others Other

Table 6: Mapping from the coarse furniture categories provided by 3D-FRONT to the fine-grained categories introduced for StyleForge. The coarse categories define the fixed furniture slots, while the fine-grained categories are used for style-prior extraction and candidate retrieval.

For each slot, Qwen selects a fine-grained category from the restricted set associated with its dataset-provided coarse category. The predicted category is accepted only if it belongs to this set. The candidate asset library is annotated and indexed using the same fine-grained taxonomy, so the validated category directly determines the retrieval pool for that slot. This design preserves the original dataset category constraints while providing StyleForge with more semantically specific slot descriptions and candidate sets.

B. Structured Style-Prior Extraction

StyleForge uses a frozen Qwen3-VL model to convert the target-style description and fixed layout into structured room- and slot-level style priors. The fixed layout is represented as a JSON object that specifies the room walls and the category, position, orientation, and size of every furniture slot. Geometry helps the model infer spatial relationships and the visual function of each slot, but no geometric attributes are included in the asset-retrieval queries.

For each room, Qwen3-VL first produces a shared room-level description covering the dominant palette, materials and finishes, and form language. Under this shared context, it then generates a category-specific query for every furniture slot. Each query must contain the original fine-grained category name, one style descriptor, one or two color descriptors, one material or finish, and one form descriptor. This formulation preserves room-level consistency while adapting the shared style to individual furniture categories. Prompt B1 gives the complete instruction.

Prompt B1: Structured style-prior extraction with Qwen3-VL.

room_id

room_type

slot_id

room_id

Room

type

room_type

Target

style

description

target_style_text

Fixed

layout

fixed_layout_json

SCHEMA

room_id

exact

input

room_id

room_type

exact

input

room_type

room_style_text

room_constraints

form_language

slot_id

exact

input

slot_id

category

exact

input

category

retrieval_text

slot_style_text

room_style_text

room_constraints

form_language

retrieval_text

retrieval_text

retrieval_text

retrieval_text

retrieval_text

slot_id

room_id

slot_style_text

slot_id

slot_id

The resulting room_style_text and room_constraints form the room-level prior that represents the shared palette, materials, and form language. For each slot, retrieval_text and slot_style_text form a category-specific prior, with retrieval_text also serving as the candidate-asset query. The room- and slot-level descriptions are subsequently embedded as and , respectively, and condition the dynamic hypergraph style field.

C. Scene-Level Evaluation Protocol

To evaluate the overall quality of the final furniture compositions, we assemble the assets selected by each method using the prescribed categories, positions, orientations, and scales, and render the complete scenes under identical camera, lighting, and rendering settings. GPT-4o receives the target-style description and the corresponding rendered scene.

GPT-4o assigns an integer score from 1 to 5 along four complementary dimensions. Aesthetic (AES) measures overall visual quality, compositional balance, and atmosphere. Color & Material (C&M) measures the coordination of colors, textures, materials, and finishes across furniture. Style Coherence (SC) measures both agreement with the target-style description and internal consistency among the selected assets. Realism & Geometry (R&G) measures visual realism, relative-scale compatibility, and geometric plausibility. Prompt C1 gives the complete evaluation instruction.

Prompt C1: GPT-4o scene-level evaluation.

target_style_text