Eliminating Rasterization: Direct Vector Floor Plan Generation with DiffPlanner

Shidong Wang and Renato Pajarola

IEEE Transactions on Visualization and Computer Graphics, 31(10): 7906-7922, 2025. (TVCG)

Getting Started

Configuration

• Python 3.9.18
• Pytorch 2.0.0
• NumPy 1.21.5

Dataset

• (Optional) Download the preprocessed RPLAN dataset from here and place it in the folder \dataset\dataset_mat. Then, run python data_preparation.py to generate the final dataset in the folder \dataset\dataset_json.
• The dataset used in this paper can also be downloaded from here. After downloading, place the dataset in the folder \dataset\dataset_json.

Training

• We provide separate training scripts (scripts\train.py) for the three sub-models (NodeDiff, AdjacencyDiff, & PartitioningDiff). You can adjust the parameters to determine which conditions each model supports. For example:

    cd node_diff/scripts
    python train.py --dataset rplan --batch_size 1024 --set_name train --support_boundary True  --support_conditions '' --support_partial False

    cd adjacency_diff/scripts
    python train.py --dataset rplan --batch_size 1024 --set_name train --support_boundary True  --support_conditions 'ncsl' --support_partial False
    
    cd partitioning_diff/scripts
    python train.py --dataset rplan --batch_size 1024 --set_name train --support_boundary True  --support_conditions 'ncsla' --support_partial False

• The trained model can be download here for testing.

Testing

• Move the trained models into the folder scripts/trained_model.
• Run the sampling scripts (scripts\sample.py) for the three sub-models in sequence, and optionally adjust the parameters to specify the input conditions. For example:

Generating bubble diagrams and floor plans using only the input boundary with entrance, without requiring any additional conditions:

    cd node_diff/scripts
    python sample.py --dataset rplan --batch_size 1024 --set_name test --model_path trained_model/b_model300000.pt --num_samples 12002 --support_boundary True --support_conditions '' --support_partial False

    cd adjacency_diff/scripts
    python sample.py --dataset rplan --batch_size 1024 --set_name test --model_path trained_model/bncsl_model300000.pt --num_samples 12002 --support_boundary True --support_conditions 'ncsl' --support_partial False --syn_dataset_path ../../output/output_json/b.json

    cd partitioning_diff/scripts
    python sample.py --dataset rplan --batch_size 1024 --set_name test --model_path trained_model/bncsla_model300000.pt --num_samples 12002 --support_boundary True --support_conditions 'ncsla' --support_partial False --syn_dataset_path ../../output/output_json/b.json

• Navigate to /output, and run python post_processing.py to get the aligned floor plans.
• Run python visualization.py, and then you can obtain the rasterized color bubble diagrams & floor plans in folder output/output_vis.

Acknowledgement

• Original RPLAN dataset: http://staff.ustc.edu.cn/~fuxm/projects/DeepLayout/index.html
• Preprocessed RPLAN dataset: https://github.com/HanHan55/Graph2plan

Citation (Bibtex)

Please cite our paper if you find it useful:

@ARTICLE{wang2025diffplanner,
  author = {Wang, Shidong and Pajarola, Renato},
  journal = {IEEE Transactions on Visualization and Computer Graphics}, 
  title = {Eliminating Rasterization: Direct Vector Floor Plan Generation with DiffPlanner}, 
  year = {2025},
  volume = {31},
  number = {10},
  pages = {7906-7922},
  doi = {10.1109/TVCG.2025.3559682}
  }