Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
RoomEditor: High-Fidelity Furniture Synthesis with Parameter-Sharing U-Net
Authors: Zhenyi Lin, Xiaofan Ming, Qilong Wang, Dongwei Ren, Wangmeng Zuo, Qinghua Hu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show our Room Editor is superior to state-of-the-arts, while generalizing directly to diverse objects synthesis in unseen scenes without task-specific fine-tuning. Our dataset and code are available at https://github.com/stonecutter-21/roomeditor. ... 5 Experiments 5.1 Experiment Setup Objective Evaluation Metrics. To compare with other methods, we use SSIM [48] and PSNR [21] for assessing reconstruction quality, FID [20] and LPIPS [58] for assessing perceptual realism, CLIP-score [38] and DINO-score [35] for assessing semantic consistency. Human Evaluation Metrics. To complement objective metrics, we conducted a user study following [7], evaluating results in fidelity (retaining reference identity and details), harmony (seamless integration with the background), and quality (overall visual appeal and detail). 5.2 Evaluation on Room Bench Our Room Editor is compared with PBE [53], Anydoor [6] and Mimicbrush [7]. These competing methods are evaluated under three settings: the released models trained on their respective large datasets, the fine-tuned models (denoted with ), whose released model weights are further fine-tuned |
| Researcher Affiliation | Academia | Zhenyi Lin Tianjin University EMAIL Xiaofan Ming Tianjin University EMAIL Qilong Wang Tianjin University EMAIL Dongwei Ren Tianjin University EMAIL Wangmeng Zuo Harbin Institute of Technology EMAIL Qinghua Hu Tianjin University EMAIL |
| Pseudocode | No | The paper describes the architecture of Room Editor using equations and textual descriptions in Section 4.1 'Room Editor Architecture' and Section 4.2 'Discussion on Merit of Room Editor', but it does not present any explicitly labeled pseudocode or algorithm blocks with structured steps. |
| Open Source Code | Yes | Our dataset and code are available at https://github.com/stonecutter-21/roomeditor. |
| Open Datasets | Yes | To address this issue, we introduce the Room Bench dataset, a ready-to-use, publicly available benchmark tailored for virtual furniture synthesis. Our Room Bench consists of 7,298 training pairs and 895 testing samples across 27 furniture categories, meticulously curated from diverse online sources. |
| Dataset Splits | Yes | To address this issue, we introduce the Room Bench dataset, a ready-to-use, publicly available benchmark tailored for virtual furniture synthesis, comprising 7,298 training pairs and 895 testing samples across 27 furniture categories. ... The training set comprises 5,288 reference furniture images and 4,094 background images, with each reference image paired with one or two background images, resulting in a total of 7,298 furniture background pairs. The testing set contains 660 reference furniture images and 895 background images, resulting in 895 testing samples. |
| Hardware Specification | Yes | Our Room Editor is initialized with the Stable Diffusion 1.5 inpainting model [40] and trained on the Room Bench dataset for 20k steps with a batch size of 32 across four NVIDIA A6000 GPUs. |
| Software Dependencies | No | Our Room Editor is initialized with the Stable Diffusion 1.5 inpainting model [40]... For the CLIP [38] model, we employ CLIP-H as the image encoder, following [7]. ... To compute the CLIP-score and DINO-score, we use CLIP Vi T-B/32 [38] and DINO Vi T-S/16 [3]... While specific models like Stable Diffusion 1.5, CLIP-H, CLIP ViT-B/32, and DINO ViT-S/16 are mentioned, the paper does not provide a comprehensive list of software dependencies with specific version numbers for the entire software stack (e.g., Python, PyTorch, CUDA versions). |
| Experiment Setup | Yes | Our Room Editor is initialized with the Stable Diffusion 1.5 inpainting model [40] and trained on the Room Bench dataset for 20k steps with a batch size of 32 across four NVIDIA A6000 GPUs. We use the Adam W optimizer [32] with a constant learning rate of 1 10 5, and set the input resolution to 512 512. |