SceneMotifCoder

SceneMotifCoder: Example-driven Visual Program Learning for Generating 3D Object Arrangements

Hou In Ivan Tam, Hou In Derek Pun, Austin T. Wang, Angel X. Chang, Manolis Savva

3DV 2025

Page | Paper | Data

Setup Environment

We recommend using mamba to manage the environment. mamba is a drop-in replacement for conda that is significantly faster and better at solving dependencies. Run the following commands to create and activate the environment. Replace mamba with conda in the following commands if you use conda.

# Create and activate the environment
mamba env create -f environment.yml
mamba activate smc

Create a .env file in the root directory of the project and add your OpenAI API key as follows:

# Inside .env
OPENAI_API_KEY=<YOUR_API_KEY>

Download Data

Example Arrangements

Download the example arrangements here and extract the contents to the root of the project.

Assets for Retrieval

SMC retrieves 3D models from the Habitat Synthetic Scenes Dataset (HSSD). To download the dataset, accept the terms and conditions of the dataset on Hugging Face here. Then, login to Hugging Face on your machine (guide) and clone the dataset repository (~72GB) at the root of the project:

cd smc

huggingface-cli login

git lfs install
git clone https://huggingface.co/datasets/hssd/hssd-models

Lastly, download the asset metadata .csv file here and place it inside the hssd-models directory.

Directory Structure

You should now have the following directory structure:

smc
├── examples
│   ├── a_stack_of_seven_plates.glb
│   ├── ...
├── hssd-models
│   ├── semantics_objects.csv
│   ├── ...
|── ...

Learn Meta-Program from Example

Run the following command to learn a meta-program from an example arrangement:

python learn.py --file examples/a_stack_of_seven_plates.glb --desc "a stack of seven plates"

The motif program and meta-program will be saved in libraries/ under the corresponding directories.

To improve a meta-program with more examples, simply run the command again with a different example arrangement of the same motif type. SMC will automatically update the meta-program using the new example.

Generate New Arrangement

After learning a meta-program, you can use it to generate new arrangements by running the following command:

python inference.py --desc "a stack of four books"

By default, the generated arrangement will be saved under outputs/. See inference.py for more options.

Create Example Arrangements

You can create your own example arrangements using your favorite 3D modeling software. Here we provide a simple guide for using Blender 3.6 LTS:

1. Create an empty object that serves as the root node (world) of the arrangement.
2. Add object meshes as children of the world node.
3. Adjust the transformations of the object meshes to create the desired arrangement.
4. Add a custom property to each object mesh. Change the property's name to semantics and type to python.
5. Set the property's value to specify the object's label. For example, {'label': 'plate'}.
6. Export the world node along with its children as a .glb file.
7. Include custom properties in the export settings by checking the Custom Properties option under Include > Data. (Default is unchecked).
8. Make sure to check the +Y Up option under Transform to ensure the correct orientation. (Default is checked).
9. Save the file.

Your new arrangement, a .glb file, is now ready to be used as an example for learning a meta-program.

Citation

Please cite our work if you find it helpful:

@article{tam2024scenemotifcoder,
    author        = {Tam, Hou In Ivan and Pun, Hou In Derek and Wang, Austin T. and Chang, Angel X. and Savva, Manolis},
    title         = {{SceneMotifCoder: Example-driven Visual Program Learning for Generating 3D Object Arrangements}},
    year          = {2024},
    eprint        = {2408.02211},
    archivePrefix = {arXiv}
}

Acknowledgements

This work was funded in part by a CIFAR AI Chair, a Canada Research Chair, NSERC Discovery Grant, and enabled by support from WestGrid and the Digital Research Alliance of Canada. We thank Qirui Wu, Jiayi Liu, and Han-Hung Lee for helpful discussions and feedback.