FreeScene

Official implementation of FreeScene: Mixed Graph Diffusion for 3D Scene Synthesis from Free Prompts (CVPR 2025).

Paper: CVPR OpenAccess | Project Page

cd FreeScene
bash settings/setup.sh
pip install -e .
python3 download_pretrain_vq.py

Dataset: we use the preprocessed data from InstructScene. Run python3 dataset/download.py (see dataset/README.md).

bash scripts/train.sh 0

Checkpoints: out/mgdit/checkpoints/

bash scripts/inference.sh 0 1099 scratch   # text-to-scene
bash scripts/inference.sh 0 1099 graph     # graph-to-scene

Set OPENAI_API_KEY (and optionally OPENAI_BASE_URL), then:

export OPENAI_API_KEY=your_key

bash scripts/generate_freescene.sh 0 1099 \
  --text "there is a sofa on the left of the double_bed" \
  --image path/to/reference.jpg

Output: out/mgdit/freescene_outputs/

@InProceedings{Bai_2025_CVPR,
  author    = {Bai, Tongyuan and Bai, Wangyuanfan and Chen, Dong and Wu, Tieru and Li, Manyi and Ma, Rui},
  title     = {FreeScene: Mixed Graph Diffusion for 3D Scene Synthesis from Free Prompts},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2025},
  pages     = {5893--5903}
}

We thank InstructScene for providing the preprocessed dataset, fVQ-VAE weights, and codebase that this project builds upon.