RoomEditor++: A Parameter-Sharing Diffusion Architecture for High-Fidelity Furniture Synthesis

Qilong Wang, Xiaofan Ming, Zhenyi Lin, Jinwen Li, Dongwei Ren, Wangmeng Zuo, Qinghua HuarXiv 2025 · DOI · arXiv

Introduction

Virtual furniture synthesis—the seamless placement of a specific object into a target room photograph—is a critical technology for modern e-commerce and interior design. Despite its potential, the task remains challenging due to the "uncanny valley" effect: even minor discrepancies in perspective, lighting, or texture alignment can erode user trust. Current solutions often struggle with feature misalignment between the reference object and the background scene, leading to visual artifacts and distorted geometries.

RoomEditor++ addresses these challenges by introducing a novel parameter-sharing diffusion architecture. By treating synthesis as a unified feature-alignment problem rather than a simple image-composition task, the model achieves state-of-the-art realism and geometric coherence.

Architecture: Parameter-Sharing Diffusion

The core innovation of RoomEditor++ is its parameter-sharing dual diffusion backbone. Traditional methods typically employ a "dual-branch" approach, using separate encoders for the object and the scene. Because these branches often utilize different architectures or weights, the extracted features fail to communicate effectively.

In contrast, RoomEditor++ uses the same diffusion model (either a U-Net or a Diffusion Transformer like FLUX.1) to process both the reference object and the background. This strategy enforces a Unified Feature Space, ensuring that both elements are represented in the same mathematical context. Reference features are then injected into the background features via a mixture attention mechanism, allowing the model to preserve the object's identity while naturally adapting its geometry to the target environment.

The RoomBench++ Dataset

To overcome the scarcity of high-quality training data, the authors introduced RoomBench++, a massive upgrade to previous benchmarks. It contains over 112,000 training pairs categorized into two distinct subsets:

  • Realistic-scene subset: Derived from professional home design renderings to ensure high-quality style coherence.
  • Real-scene subset: Extracted from heterogeneous real-world indoor videos, capturing natural lighting, occlusions, and the "noisy" traits of real-world photography.
Comparison of Realistic-scene data (professional renderings) and Real-scene data (noisy real-world video) used in RoomBench++.

Key Results and Generalization

RoomEditor++ significantly outperforms existing state-of-the-art methods such as AnyDoor and MimicBrush. Quantitatively, the DiT-based version achieved a FID of 11.49 on the RoomBench++ test set, nearly doubling the performance of its closest competitors. Qualitative user studies further confirmed these findings, with RoomEditor++ being preferred in 33.2% of cases for overall quality.

Qualitative results showing RoomEditor++'s ability to synthesize diverse objects—from furniture to toys—with high geometric and semantic fidelity.

Beyond furniture, the model demonstrates impressive zero-shot generalization. Even when trained primarily on interior decor, it successfully integrates diverse objects like backpacks, stuffed animals, and candles into scenes with high fidelity, proving the robustness of its feature-sharing logic.

Conclusion and Future Work

By unifying the feature space through parameter sharing and leveraging the massive RoomBench++ dataset, RoomEditor++ sets a new standard for virtual furniture synthesis. While the model implicitly learns many environmental cues, future work aims to integrate explicit physical constraints—such as complex light-source interactions and ray-traced shadows—to further enhance physical plausibility for professional design applications.

Summary updated

Abstract

Virtual furniture synthesis, which seamlessly integrates reference objects into indoor scenes while maintaining geometric coherence and visual realism, holds substantial promise for home design and e-commerce applications. However, this field remains underexplored due to the scarcity of reproducible benchmarks and the limitations of existing image composition methods in achieving high-fidelity furniture synthesis while preserving background integrity. To overcome these challenges, we first present RoomBench++, a comprehensive and publicly available benchmark dataset tailored for this task. It consists of 112,851 training pairs and 1,832 testing pairs drawn from both real-world indoor videos and realistic home design renderings, thereby supporting robust training and evaluation under practical conditions. Then, we propose RoomEditor++, a versatile diffusion-based architecture featuring a parameter-sharing dual diffusion backbone, which is compatible with both U-Net and DiT architectures. This design unifies the feature extraction and inpainting processes for reference and background images. Our in-depth analysis reveals that the parameter-sharing mechanism enforces aligned feature representations, facilitating precise geometric transformations, texture preservation, and seamless integration. Extensive experiments validate that RoomEditor++ is superior over state-of-the-art approaches in terms of quantitative metrics, qualitative assessments, and human preference studies, while highlighting its strong generalization to unseen indoor scenes and general scenes without task-specific fine-tuning. The dataset and source code are available at \url{https://github.com/stonecutter-21/roomeditor}.
BibTeX
@misc{wang2025roomeditor,
  author = {Qilong Wang and Xiaofan Ming and Zhenyi Lin and Jinwen Li and Dongwei Ren and Wangmeng Zuo and Qinghua Hu},
  title = {{RoomEditor++: A Parameter-Sharing Diffusion Architecture for High-Fidelity Furniture Synthesis}},
  year = {2025},
  eprint = {2512.17573},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.17573},
}
Figure