[ICLR 2025] EditRoom: LLM-parameterized Graph Diffusion for Composable 3D Room Layout Editing

Kaizhi Zheng, Xiaotong Chen, Xuehai He, Jing Gu, Linjie Li, Zhengyuan Yang, Kevin Lin, Jianfeng Wang, Lijuan Wang, Xin Eric Wang

<p> <img width="730" alt="pipeline", src="./assets/new_teaser-1.png"> </p> </h4>

This repository contains the official implementation of the paper: EditRoom: LLM-parameterized Graph Diffusion for Composable 3D Room Layout Editing, which is accepted by ICLR 2025. EditRoom is a graph diffusion-based generative model, which can manipulate 3D Room by using natural lanauge.

Feel free to contact me (kzheng31@ucsc.edu) or open an issue if you have any questions or suggestions.

Getting Start

1. Installation

Clone our repo and create a new python environment.

git clone https://github.com/eric-ai-lab/EditRoom.git
cd EditRoom
conda create -n editroom python=3.10
conda activate editroom
pip install -r requirements.txt

Download the Blender software for visualization.

cd blender
wget https://download.blender.org/release/Blender3.3/blender-3.3.1-linux-x64.tar.xz
tar -xvf blender-3.3.1-linux-x64.tar.xz
rm blender-3.3.1-linux-x64.tar.xz

2. Creating Dataset

Dataset used in EditRoom is based on 3D-FORNT and 3D-FUTURE. Please refer to the instructions provided in their official website to download the original dataset. Based on InstructScene, we write an automatic dataset generator to generate scene editing pairs. We provided preprocessed datasets on HuggingFace for start.

First, downloading the preprocessed datasets. They will be downloaded under datasets folder by default. If you want to change to another directory, please remind to chage environment viariable EDITROOM_DATA_FOLDER.

export EDITROOM_DATA_FOLDER="./datasets"
python3 tools/download_dataset.py

Then, please refer to tools/README.md for more details.

3. Training

We will first train scene graph to scene layout generator:

python3 src/train_edit.py --config_file configs/bedroom_sg2sc_diffusion.yaml --output_directory ./weights --with_wandb_logger

Then, we will train scene graph generator:

python3 src/train_edit.py --config_file configs/bedroom_sg_diffusion.yaml --output_directory ./weights --with_wandb_logger

You can change config_file to other room types. All config files are under configs folder.

4. Evaluation

To run the evaluation, we need both SG_WEIGHT and SG2SC_WEIGHT. Those weights should be found under output_directory for training.

If you runing on a headless server, please using Xvfb to create virtual screens for pyrender.

tmux new -s v_screen #Opening another terminal at the backend
sudo Xvfb :99 -screen 0 1960x1024x24

For evaluation, please run:

export DISPLAY=":99" #For headless server
python3 --sg_config_file configs/bedroom_sg_diffusion.yaml \
        --sg2sc_config_file configs/bedroom_sg2sc_diffusion.yaml \
        --output_directory ./weights \
        --sg_weight_file SG_WEIGHT \
        --sg2sc_weight_file SG2SC_WEIGHT \
        --llm_plan_path LLM_PLAN

LLM_PLAN should be created during dataset generation. Please refer to tools/README.md for more details.

You can also using tempalte commands for evaluation.

export DISPLAY=":99" #For headless server
python3 --sg_config_file configs/bedroom_sg_diffusion.yaml \
        --sg2sc_config_file configs/bedroom_sg2sc_diffusion.yaml \
        --output_directory ./weights \
        --sg_weight_file SG_WEIGHT \
        --sg2sc_weight_file SG2SC_WEIGHT

Acknowledgement

We would like to thank the authors of ATISS, DiffuScene, OpenShape, NAP, CLIPLayout and InstructScene for their great work and generously providing source codes, which inspired our work and helped us a lot in the implementation.

Citation

If you find our work helpful, please consider citing:

@inproceedings{
zheng2025editroom,
title={EditRoom: {LLM}-parameterized Graph Diffusion for Composable 3D Room Layout Editing},
author={Kaizhi Zheng and Xiaotong Chen and Xuehai He and Jing Gu and Linjie Li and Zhengyuan Yang and Kevin Lin and Jianfeng Wang and Lijuan Wang and Xin Eric Wang},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025}
}