Paper | arXiv | Video | Project Page

This is the repository that contains source code for the paper:

DiffuScene: Scene Graph Denoising Diffusion Probabilistic Model for Generative Indoor Scene Synthesis

• We present DiffuScene, a diffusion model for diverse and realistic indoor scene synthesis.

If you find DiffuScene useful for your work please cite:

@inproceedings{
    tang2023diffuscene,
    title={DiffuScene: Scene Graph Denoising Diffusion Probabilistic Model for
      Generative Indoor Scene Synthesis},
    author={Tang, Jiapeng and Nie Yinyu and Markhasin Lev and Dai Angela and Thies Justus and Nie{\ss}ner, Matthias},
    booktitle={arxiv},
    year={2023},
    }

Contact Jiapeng Tang for questions, comments and reporting bugs.

You can create a conda environment called diffuscene using

conda env create -f environment.yaml
conda activate diffuscene

Next compile the extension modules. You can do this via

python setup.py build_ext --inplace
pip install -e .

Install ChamferDistancePytorch

cd ChamferDistancePytorch/chamfer3D
python setup.py install

The training and evaluation are based on the 3D-FRONT and the 3D-FUTURE dataset. To download both datasets, please refer to the instructions provided in the dataset's webpage.

To accelerate the preprocessing speed, we can sepcify the PATH_TO_SCENES environment variable for all scripts. This filepath contains the parsed ThreedFutureDataset after being pickled. To pickle it, you can simply run this script as follows:

python pickle_threed_future_dataset.py path_to_output_dir path_to_3d_front_dataset_dir path_to_3d_future_dataset_dir path_to_3d_future_model_info --dataset_filtering room_type

Based on the pickled ThreedFutureDataset, we also provide a script to pickle the sampled point clouds of object CAD models, which are used to shape autoencoder training and latent shape code extraction.

python pickle_threed_future_pointcloud.py path_to_output_dir path_to_3d_front_dataset_dir path_to_3d_future_dataset_dir path_to_3d_future_model_info --dataset_filtering room_type

For example,

python pickle_threed_future_dataset.py  /cluster/jtang/3d_front_processed/ /cluster/jtang/3D-FRONT/ /cluster/jtang/3D-FUTURE-model /cluster/balrog/jtang/3D-FUTURE-model/model_info.json  --dataset_filtering threed_front_livingroom --annotation_file ../config/livingroom_threed_front_splits.csv

PATH_TO_SCENES="/cluster/jtang/3d_front_processed/threed_front.pkl" python pickle_threed_fucture_pointcloud.py /cluster/jtang/3d_front_processed/ /cluster/jtang/3D-FRONT/ /cluster/jtang/3D-FUTURE-model /cluster/jtang/3D-FUTURE-model/model_info.json  --dataset_filtering threed_front_livingroom --annotation_file ../config/livingroom_threed_front_splits.csv

Note that these two scripts should be separately executed for different room types containing different objects. For the case of 3D-FRONT this is for the bedrooms and the living/dining rooms, thus you have to run this script twice with different --dataset_filtering and --annotation_fileoptions. Please check the help menu for additional details.

Then you can train the shape autoencoder using all models from bedrooms/diningrooms/livingrooms.

cd ./scripts
PATH_TO_SCENES="/cluster/jtang/3d_front_processed/threed_front.pkl" python train_objautoencoder.py ../config/obj_autoencoder/bed_living_diningrooms_lat32.yaml your_objae_output_directory --experiment_tag  "bed_living_diningrooms_lat32" --with_wandb_logger

Next, you can use the pre-train checkpoint of shape autoencoder to extract latent shape codes for each room type. Take the bedrooms for example:

PATH_TO_SCENES="/cluster/jtang/3d_front_processed/threed_front.pkl" python generate_objautoencoder.py ../config/objautoencoder/bedrooms.yaml your_objae_output_directory --experiment_tag "bed_living_diningrooms_lat32"

Finally, you can run preprocessing_data.py to read and pickle object properties (class label, location, orientation, size, and latent shape features) of each scene.

PATH_TO_SCENES="/cluster/jtang/3d_front_processed/threed_front.pkl" python preprocess_data.py /cluster/jtang/3d_front_processed/livingrooms_objfeats_32_64 /cluster/jtang/3D-FRONT/ /cluster/jtang/3D-FUTURE-model /cluster/jtang/3D-FUTURE-model/model_info.json --dataset_filtering threed_front_livingroom --annotation_file ../config/livingroom_threed_front_splits.csv --add_objfeats

The proprossed datasets can also be downloaded from here.

To train diffuscene on 3D Front-bedrooms, you can run

./run/train.sh

To generate the scene of unconditional diffusion model, you can run

./run/generate.sh

Please also check out the following papers that explore similar ideas:

• Fast and Flexible Indoor Scene Synthesis via Deep Convolutional Generative Models pdf
• Sceneformer: Indoor Scene Generation with Transformers pdf
• ATISS: Autoregressive Transformers for Indoor Scene Synthesis pdf
• Indoor Scene Generation from a Collection of Semantic-Segmented Depth Images pdf
• Scene Synthesis via Uncertainty-Driven Attribute Synchronization pdf
• LEGO-Net: Learning Regular Rearrangements of Objects in Roomspdf

Most of the code is borrowed from ATISS. We thank for Despoina Paschalidou her great works and repos.