Initial code release for the Sceneformer paper, contains models, train and test scripts for the shape conditioned model. Text conditioned model and detailed README coming soon.

Please also check the project website here


Install the requirements in requirements.txt and environment.yaml in a conda environment. Packages that are common can be installed either through
pip or conda.

Prepare Data

The SUNCG dataset is currently not available, hence all related files have been removed. The dataset can be prepared with the scripts which were taken from deepsynth.


Configure the experiment in configs/scene_shift_X_config.yaml where X is one of cat, dim, loc, ori

Then run

python scene_scripts/ configs/scene_shift_X_config.yaml

to train the model X.


Configure the model paths in scene_scripts/ and then run

python scene_scripts/

If you find our work useful, please consider citing us:

  title={SceneFormer: Indoor Scene Generation with Transformers},
  author={Wang, Xinpeng and Yeshwanth, Chandan and Nie{\ss}ner, Matthias},
  journal={arXiv preprint arXiv:2012.09793},