Environment

Preparation

To prepare data and pretraind models, please check all the file folders in this project and follow the guidance in readme.md.
To accelerate the training process, we precompute the 512-dim style code for every artistic human face, you can download style_codes.csv and place it under ArtNerf/.

The model is trained by conducting a two-stage training strategy: pretraining on CelebA and fine-tuning on both AAHQ and CelebA.
The whole model is composed of 1 generator and 3 dicriminators. disc_real guides the gen to generate natural human faces and disc_style guides the gen to generate stylized human faces. disc_latent helps ensure the style-consistency between generated faces and ref faces.
We use a style blending module to help stabilize the cross-domain transfer learning process and allow users to change the extent to which the generated images is stylized(level can be changed from 0 to 11).

Following the traditional face synthesis models like StyleGAN, we can perform interpolation between any two latent codes.

| | | | | | | | | | | |

Name		Name	Last commit message	Last commit date
Latest commit History 33 Commits
LPIPS		LPIPS
ckpt		ckpt
data		data
discriminators		discriminators
experiments		experiments
generators		generators
models		models
pretrained		pretrained
ref_imgs		ref_imgs
curriculums.py		curriculums.py
datasets.py		datasets.py
fid_evaluation.py		fid_evaluation.py
readme.md		readme.md
train_artnerf.py		train_artnerf.py