Time of the Flight of the Gaussians: Optimizing Depth Indirectly in Dynamic Radiance Fields

Project Page | Paper | arXiv (Latest) | Pre-trained Models | F-TöRF Data | TöRF Data

Runfeng Li, Mikhail Okunev, Zixuan Guo, Anh Ha Duong, Christian Richardt^∞, Matthew O'Toole^*, James Tompkin
Brown University, ^∞Meta Reality Labs, *Carnegie Mellon University

Quick Start

To set up the environment, run

conda env create -f environment.yml
conda activate gftorf

If your machine (hardware & software) is compatible with the original 3DGS, then you should have no problem setting up our environment.

Next, you have two options:

Option A: Render using pre-trained models

1. To download pretrained models, run:

   python prepare_models.py
   

2. Modify arguments in run_render.py if needed, then run:
   (You might need to modify the IMAGEMAGICK path in conf.py to compose video panels.)

   python run_render.py
   

   You should get the exact video panels as shown on our project page, such as:

   

Option B: Optimize from scratch

This can take at most 60 minutes for one scene on a single NVIDIA 3090 GPU.

1. Download F-TöRF real_scenes.zip, synthetic_scenes.zip and TöRF copier, cupboard, deskbox, phonebooth, and studbook scenes to the data/ folder, and then run:

   python prepare_data.py
   

2. Modify arguments in run_optimize.py if needed, then run:

   python run_optimize.py
   

   You can get some decent looking results after 20k iterations, such as:
   (though training longer would usually be better)

   

Citation

@InProceedings{Li_2025_CVPR,
  author    = {Li, Runfeng and Okunev, Mikhail and Guo, Zixuan and Duong, Anh Ha and Richardt, Christian 
              and O'Toole, Matthew and Tompkin, James},
  title     = {Time of the Flight of the Gaussians: Optimizing Depth Indirectly in Dynamic Radiance Fields},
  booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
  month     = {June},
  year      = {2025},
  pages     = {21021-21030}
}