Text-Guided Synthesis of Crowd Animation (SIGGRAPH 2024)

Code for SIGGRAPH 2024 paper "Text-Guided Synthesis of Crowd Animation".

Preparation

1. Install packages in requirements.txt. The implementation is based on Pytorch.
2. Find package pyDeclutter and RVO2_Python in Libs. For each library, run python setup.py build to build, and python setup.py install to install.
3. Visit Diffusers to install the Diffusers, which is a modular library that contains most of the SOTA pre-trained diffusion models.

Dataset

You can download the already generated and post-processed dataset from this link. Download Dataset.zip to get the dataset and unzip it into the Language_Crowd_Animation folder for use.

Or you can generate it by yourself:

1. Run Dataset_Generation.py to generate the initial dataset. The generated dataset contains velocity fields without optimization, which would push the agents to be concentrated.
2. Run Dataset_Postprocess.py to post-process the initial dataset.

Training

You can directly use the pre-trained diffusion models from this link. Download Models_Server_ForTest.zip to get the pre-trained diffusion models and unzip it into the Language_Crowd_Animation folder for use.

Or you can train the models by yourself:

1. Run Trainer_SgDistrDiffusion_Full_V1_Server.py to train the start and goal diffusion model.
2. Run Trainer_FieldDiffusion_Full_V2_Server.py to train the velocity field diffusion model.

Inference

Run Quantitative_Exps.py to evaluate the model using testing data from dataset.

Citation

Please consider citing our paper if our code is useful for your research:

@inproceedings{ji2024text,
  title={Text-Guided Synthesis of Crowd Animation},
  author={Ji, Xuebo and Pan, Zherong and Gao, Xifeng and Pan, Jia},
  booktitle={ACM SIGGRAPH 2024 Conference Papers},
  pages={1--11},
  year={2024}
}

Contact

Feel free to email [email protected] or [email protected] if you have any questions.