RoomBench++
收藏资源简介:
RoomBench++是由天津大学与哈尔滨工业大学联合构建的开放式家具合成基准数据集,包含专业家居设计渲染图(7,298训练对)和真实室内视频帧(105,553训练对)两大子集,总计114,683对数据。该数据集通过自动化流程采集并标注,覆盖多样化家具类别与真实场景特征,旨在解决虚拟家具合成领域缺乏实用基准的难题,为高保真家具融入室内场景的算法研发提供评估基础,广泛应用于家居设计、电子商务等场景的视觉合成任务。
RoomBench++ is an open furniture synthesis benchmark dataset jointly developed by Tianjin University and Harbin Institute of Technology. It consists of two subsets: professional home design renderings (7,298 training pairs) and real indoor video frames (105,553 training pairs), with a total of 114,683 data pairs in all. Collected and annotated via automated workflows, the dataset covers diverse furniture categories and realistic indoor scene characteristics. It aims to address the critical shortage of practical benchmarks in the field of virtual furniture synthesis, providing a solid evaluation foundation for the research and development of algorithms that integrate high-fidelity furniture into indoor scenes, and is widely utilized for visual synthesis tasks in scenarios such as home design and e-commerce.
RoomEditor数据集概述
数据集名称
RoomEditor: High-Fidelity Furniture Synthesis with Parameter-Sharing U-Net
关联会议
NeurIPS 2025
核心内容
- 发布RoomEditor模型。
- 发布RoomBench数据集。
- 发布评估代码。
发布日期
2025年12月15日
模型与数据下载
- SD-1.5-inpainting检查点:可从HuggingFace的stable-diffusion-inpainting下载。
- RoomEditor模型权重与测试图像:可从https://share.multcloud.link/share/74f20132-4269-41d0-8355-0c39106de6b0下载。
评估与演示
- 运行推理评估:使用命令
bash run.sh。 - 运行Gradio演示:使用脚本
python run_gradio_demo.py。
环境依赖
安装要求:pip install -r requirements.txt

- 1RoomEditor++: A Parameter-Sharing Diffusion Architecture for High-Fidelity Furniture Synthesis天津大学人工智能学院, 哈尔滨工业大学计算机科学与技术学院 · 2025年



