Environment

Preparation

Place 128x128 natural huaman faces(around 200,000) from CelebA under data/celeba/.
Place the first 23567 128x128 natural huaman faces from CelebA under data/celeba_mini, which can be downloaded from here.
Place 23567 128x128 artistic huaman faces from processed AAHQ dataset under data/aahq/.
You can also replace CelebA with other datasets like FFHQ to achieve higher visual quality. Note that if you do this, you should change configs related to resolution(img_size and img_size_sr, etc.)
Pretrained models
Download and unzip the file and put style_encoder.pt under ckpt/
Download and unzip the file. Place all the related file folders (artnerf_models, base_models) under experiments/
Place vgg19-dcbb9e9d.pth (which can be downloaded from PyTorch official website) in this directory.
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).