SceneScape: Text-Driven Consistent Scene Generation

[Project Page]

cave_haunted.mp4

SceneScape is a method for generating videos of indoor scenes walkthroughs from text, as described here.

Abstract

We present a method for text-driven perpetual view generation -- synthesizing long-term videos of various scenes solely, given an input text prompt describing the scene and camera poses. We introduce a novel framework that generates such videos in an online fashion by combining the generative power of a pre-trained text-to-image model with the geometric priors learned by a pre-trained monocular depth prediction model. To tackle the pivotal challenge of achieving 3D consistency, i.e., synthesizing videos that depict geometrically-plausible scenes, we deploy an online test-time training to encourage the predicted depth map of the current frame to be geometrically consistent with the synthesized scene. The depth maps are used to construct a unified mesh representation of the scene, which is progressively constructed along the video generation process. In contrast to previous works, which are applicable only to limited domains, our method generates diverse scenes, such as walkthroughs in spaceships, caves, or ice castles.

Getting Started

Installation

For the installation to be done correctly, please proceed only with CUDA-compatible GPU available.

Clone the repo and create the environment:

git clone https://github.com/rafailfridman/SceneScape.git
cd SceneScape
conda create --name scenescape python=3.10
conda activate scenescape

We are using Pytorch3D to perform rendering. Run the following commands to install it or follow their installation guide (it may take some time).

conda install pytorch=1.13.0 torchvision pytorch-cuda=11.6 -c pytorch -c nvidia
conda install -c fvcore -c iopath -c conda-forge fvcore iopath
conda install -c bottler nvidiacub
conda install pytorch3d -c pytorch3d

Install the rest of the requirements:

pip install -r requirements.txt

Splatting installation (OPTIONAL)

If you want to test how our method works without the mesh, just by warping pixels from frame i to i+1, you will have to install softmax-splatting package, following their instructions. Initialize the submodule:

git submodule update --init

Then to run with splatting and not the mesh, change the use_splatting flag in the configs/base-config.yaml file to True.

Run examples

SceneScape is designed to generate walkthough videos of static scenes. It is not designed for generating dynamic content (e.g. people, cars moving, etc.).
Running SceneScape multiple times with the same inputs can lead to slightly different results.
Method is not designed to generate outdoor scenes because of the mesh used as a unified representation.

Method was tested to run on a single Tesla V100 32GB, and takes ~20GB of video memory. It takes approximately 2.5 hours to generate a 50 frames video on a single Tesla V100 32GB.

Run the following command to start training

python run.py --example_config config/example_configs/dungeon.yaml

Intermediate results will be saved to output during optimization. Once the method is done, the video will be saved to output/output.mp4. You can change the prompt in the config/example_configs/dungeon.yaml file to generate different videos, or create your own config. The base parameters of our method are located in configs/base-config.yaml.

For more results see the supplementary material.

Citation

@article{SceneScape,
      author    = {Rafail Fridman and Amit Abecasis and Yoni Kasten and Tali Dekel},
      title     = {SceneScape: Text-Driven Consistent Scene Generation},
      journal   = {arXiv preprint arXiv:2302.01133},
      year      = {2023},
  }

rimless/SceneScape