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
cs.CV
2026-08-03
Per-piece furniture retrieval hits only 22.8 R@1; StyleForge's whole-room hypergraph with counterfactual Mahalanobis lifts it to 79.1, scoring 4.59 to nearly match human pairings.
Given a fixed room layout (every furniture piece's category, position, orientation, and scale already set), the task is to choose a concrete 3D asset for each slot so that the whole room reads as coherent. The catch is that a single piece carries only partial style semantics; real room style emerges from spatial, functional, and global relations among several pieces. Existing methods mostly retrieve each slot independently (ULIP, OpenShape, Uni3D score each asset with CLIP-style representations), so individually fine assets clash in shape, material, and color once composed. The few relation-based methods rely on fixed pairwise edges and cannot express higher-order dependencies that shift with layout and style. StyleForge attacks this mismatch between per-piece retrieval and whole-room coherence.
StyleForge frames furniture styling as joint optimization over all slots rather than per-piece ranking.
On 3D-FRONT (7,100 training rooms, 800 test rooms), with GPT-4o scoring aesthetics (AES), color and material (C&M), style coherence (SC), and realism and geometry (R&G) on a 1 to 5 scale:
| Method | Final R@1 | Avg score |
| ULIP | 33.9 | 3.04 |
| OpenShape | 35.1 | 3.08 |
| Uni3D | 36.3 | 3.14 |
| MetaFind | 44.5 | 4.25 |
| StyleForge | 79.1 | 4.59 |
| Human (GT) | — | 4.63 |
The initial retrieval R@1 is only 22.8; after scene-level optimization it reaches 79.1, with an average score of 4.59 that nearly matches the human 4.63. In the ablation, retrieval prior alone stays at 22.8, adding the hypergraph and iteration barely moves it, and only adding Mahalanobis counterfactuals jumps it to 79.1. By room type, bedroom is 82.4 and dining 80.6, while the slot-heavy living room is lowest at 75.8. In a blind professional review, the high-scoring tier is approved 86.1% of the time (95% CI 71.3% to 93.9%) against 22.2% for the low tier.
For anyone working on 3D scene or asset generation, StyleForge is a clear methodological shift from per-piece retrieval to whole-room joint optimization, and it puts test-time training at inference (freeze the LLM, tune only the room's logits) where the cost is controllable. The Mahalanobis-versus-Euclidean result is worth remembering: on a structure like style, where some dimensions are strongly constrained and others are noise, a distance with a learned variance is markedly more accurate. By the numbers it pushes automatic styling to near-human level, a real step for interior-design automation.
The authors' main admitted limit: the whole pipeline depends on the initial retriever's recall. If the right asset never enters the top-K, scene-level optimization cannot recover it; they plan to use scene-level feedback to update the retrieval query dynamically. Two more concerns stand out. The evaluation leans on GPT-4o scoring; although a designer blind review backs it, R@1 as "hit the human annotation" assumes the human pairing is the unique correct answer, while furniture styling has many valid solutions. Inference also runs a TTT optimization per room (up to 200 steps), and the paper does not clarify the real-time cost.