DO-FAM

Official implementation of DO-FAM

Quickstart

Pretrained models

Please download the pretrained models and move them to ./pretrained_models/

Note: when test, there is no need to use classifier, so you don't need to download it.

Model	Description
DOLL	Pretrained models of DOLL on attribute Gender, Eyeglasses, Age, Smiling
hyperstyle+stylegan	Pretrained generator released by hyperstyle
wEncoder	Pretrained face encoder released by hyperstyle

Test datas

Please download the test dataset(30 images with corresponding latent codes and weight deltas) in this link and put them in ./test_data/

Then run the following command:

python test.py --attribute 'Smiling' --coeff_min -1.5 --coeff_max 1.5 --step 0.5 --gpu '0'

Test on own images

for lower GPU usage, we seperate the editing process to 2 steps: inversion step and editing step

Step1: run inversion to get latent codes and weights deltas

1. Download checkpoints mentioned in the upper chart.

2. Modify data_paths.test_image to your own image paths in ./configs/path_config.py.

3. Run the following command to get the latent codes and weights of images(will take 7G RAM in GPU):

python generate_latents_and_weights.py  --exp_dir './test_data/' --save_weight_deltas --gpu '0'

Step2: run editing step to get result

1. Modify data_paths.test_latent and data_paths.test_weights_delta in ./configs/path_config.py to your own in Step1

2. Run the following command to get the edited image(will take 8G RAM in GPU):

python test.py \
--test_latent_path './test_data/latent_codes/' \
--test_weights_delta_path './test_data/weight_deltas/' \
--save_image_path './test_data/' \
--attribute 'Smiling' --coeff_min -1.5 --coeff_max 1.5 --step 0.5 --gpu '0'

You can choose 'Eyeglasses', 'Gender', 'Smiling' and 'Age' to manipulate.

Training

To run the training, please download the extra pretrained models and move them to ./pretrained_models/

Model	Description
classifier	Pretrained models of Latent Classifier on hyperstyle latent codes for 40 attributes