The ability to create high-quality 3D faces from a single image has become in- creasingly important with wide applications in video conferencing, AR/VR, and advanced video editing in movie industries. In this paper, we propose Face Diffu- sion NeRF(FDNeRF), a new generative method to reconstruct high-quality Face NeRFs from single images, complete with semantic editing and relighting capabili- ties. FDNeRF utilizes high-resolution 3D GAN inversion and expertly trained 2D latent-diffusion model, allowing users to manipulate and construct Face NeRFs in zero-shot learning without the need for explicit 3D data. With carefully designed illumination and identity preserving loss, as well as multi-modal pre-training, FD- NeRF offers users unparalleled control over the editing process enabling them to create and edit face NeRFs using just single-view images, text prompts, and explicit target lighting. The advanced features of FDNeRF have been designed to produce more impressive results than existing 2D editing approaches that rely on 2D segmentation maps for editable attributes. Experiments show that our FDNeRF achieves exceptionally realistic results and unprecedented flexibility in editing compared with state-of-the-art 3D face reconstruction and editing methods

• We recommend Linux for performance and compatibility reasons.
• 1–2 high-end NVIDIA GPUs. We have done all testing and development using V100s, RTX3090s and RTX4090s.
• 64-bit Python 3.9, cuda11.3, and PyTorch 1.11.0 (or later). See https://pytorch.org for PyTorch install instructions.
• Since we use the EG3D as our backbone, Please see eg3d official repo for EG3D installation. Or directly install following conda environment.
• Python libraries: see environment.yml for exact library dependencies. You can use the following commands with Anaconda3 to create and activate your Python environment:
  • cd FaceDNeRF
  • conda env create -f environment.yml
  • conda activate facednerf

We use the same camera pose convention as eg3d, please refer to this script that can preprocess in-the-wild images compatible with our camera pose convention.

We also provide test data in the ./test_data

1. VGG16 pre-trained model: you can download vgg16.pt from https://nvlabs-fi-cdn.nvidia.com/stylegan2-ada-pytorch/pretrained/metrics/vgg16.pt and save it to ./network.
2. Since EG3D is the backbone of our model, please download the ffhqrebalanced512-64.pkl by the following command and place it under ./network.

wget --content-disposition 'https://api.ngc.nvidia.com/v2/models/org/nvidia/team/research/eg3d/1/files?redirect=true&path=ffhqrebalanced512-128.pkl' -O ffhqrebalanced512-128.pkl

3. The ArcFace facial recognition network is used in the ID loss. The weights can be downloaded from here.

The backnone of this implementation is eg3d. For detailed instructions, please refer to the comments in the ./script.py

python script.py

Results will be saved to ./output. If you encounte the error: Function 'PowBackward0' returned nan values in its 0th output. caused by the File "XXX/anaconda3/envs/facednerf/lib/python3.9/site-packages/kornia/color/rgb.py", line 199, in rgb_to_linear_rgb, please replace the line 199 code: lin_rgb: torch.Tensor = torch.where(image > 0.04045, torch.pow(((image + 0.055) / 1.055), 2.4), image / 12.92) with lin_rgb: torch.Tensor = torch.where(image > 0.04045, torch.pow(((image.abs() + 0.055) / 1.055), 2.4), image / 12.92).Since the .abs() function is added to ensure that the base number is positive, the torch.power function becomes differentiable, and the gradients can be passed through it.

we have provided some illumination functions getLightIntensityin the file ./editors/w_plus_editor.py. If you want to customize you own illumination function, pleace also name it getLightIntensity and replace the original one.

Link to Paper

If you find this work useful for your research, please cite our paper:

@article{zhang2024facednerf,
  title={FaceDNeRF: Semantics-Driven Face Reconstruction, Prompt Editing and Relighting with Diffusion Models},
  author={Zhang, Hao and DAI, Tianyuan and Xu, Yanbo and Tai, Yu-Wing and Tang, Chi-Keung},
  journal={Advances in Neural Information Processing Systems},
  volume={36},
  year={2024}
}