eckertzhang / Text2NeRF

Official implementation of 'Text2NeRF: Text-Driven 3D Scene Generation with Neural Radiance Fields'
MIT License
123 stars 6 forks source link

Text2NeRF

Official implementation of 'Text2NeRF: Text-Driven 3D Scene Generation with Neural Radiance Fields'. Note: This code is forked from TensoRF.

Installation

Install environment:

conda env create -f environment.yml
conda activate text2nerf
pip install -r requirements.txt

Download Pre-trained Weights

  1. Download pre-trained 'CLIPModel(clip-vit-base-patch32)' into the folder 'weights'.
  2. Download pre-trained 'SDM' into the folder 'weights'.
  3. Download pre-trained 'SDM-Inpaint(stable-diffusion-2-inpainting)' into the folder 'weights'.
  4. Download other required weights for third parties from Google_Drive, and put it into the folder 'weights'

Optimization

The training script is in text2nerf_main.py:

Local Scene Generation

CUDA_VISIBLE_DEVICES=0 python text2nerf_main.py --config 'configs/text2nerf_scenes.txt' --expname 'text000' --prompt 'a beautiful garden' --datadir 'data_example/text000' --pose_traj 'local_fixed' --regen_pose

360-Degree Scene Generation

CUDA_VISIBLE_DEVICES=0 python text2nerf_main.py --config 'configs/text2nerf_scenes.txt' --expname 'text000_360' --prompt 'a beautiful garden' --datadir 'data_example/text000' --pose_traj 'circle' --regen_pose

Rendering

CUDA_VISIBLE_DEVICES=0 python text2nerf_main.py --config 'configs/text2nerf_scenes.txt' --expname 'text000' --prompt 'a beautiful garden' --datadir 'data_example/text000' --pose_traj 'local_fixed' --regen_pose --render_only 1 --render_test 1 --ckpt '[path of ckpt]'

Citation

If you find our code or paper helps, please consider citing:

@article{zhang2023text2nerf,
  title={Text2NeRF: Text-Driven 3D Scene Generation with Neural Radiance Fields},
  author={Zhang, Jingbo and Li, Xiaoyu and Wan, Ziyu and Wang, Can and Liao, Jing},
  journal={arXiv preprint arXiv:2305.11588},
  year={2023}
}