Xuyang Guo, Meina Kan, Zhenliang He, Xingguang Song, Shiguang Shan

Instance-level facial attribute transfer aims at transferring an attribute including its style from a source face to a target one. Existing studies have limitations on fidelity or correctness. To address this problem, we propose a weakly supervised style disentangling method embedded in Generative Adversarial Network (GAN) for accurate instance-level attribute transfer, using only binary attribute annotations. In our method, the whole attributes transfer process is designed as two steps for easier transfer, which first removes the original attribute or transfers it to a neutral state and then adds the attributes style disentangled from a source face. Moreover, a style disentangling module is proposed to extract the attribute style of an image used in the adding step. Our method aims for accurate attribute style transfer. However, it is also capable of semantic attribute editing as a special case, which is not achievable with existing instance-level attribute transfer methods. Comprehensive experiments on CelebA Dataset show that our method can transfer the style more precisely than existing methods, with an improvement of 39% in user study, 16.5% in accuracy, and about 3.3 in FID.

Clone this repo.

git clone https://github.com/XuyangGuo/STD-GAN.git
cd STD-GAN

The code requires python 3.6.

We recommend using Anaconda to regulate packages.

Dependencies:

• PyTorch 1.4
• torchvision, tensorboardX, pillow, skimage
• tqdm, addict

Please download the pre-trained model, move it to ./ with the correct path CtrlHair/model_trained.

python test.py -a @AttributesName -t @TargetImagePath -s @SourceImagePath -r @ResultImagePath -c 001

Here are explanations of parameters:

• -a @AttributesName The edited attribute name. Make sure the attribute name appears in the use_atts field of the config in config.py. For example, the 001 config can use Bangs, Smiling, Eyeglasses, Black_Hair (Black_Hair here refers to turning black hair into other colors). Separate multi attributes to be edited with commas and without spaces.
• -t @TargetImagePath The path of target image to edit. Please make sure that the crop method of this image is same as training set. The pretrained model uses CelebA official cropped images.
• -s @SourceImagePath The path of style source image. The requirements for the image are the same as the target image. If the source image is not provided, a random style code is sampled from the uniform distribution U(-1,1).
• -r @ResultImagePath The path of saving the result image.
• -c xxx By default, the value is 001, which corresponds to the config in experiment_name in config.py. If you want to use other config, please add it to config.py and call it by the prefix of its experiment_name.

We provide some target images and sample images of each attribute in input_imgs for convenient usage. Example of usage:

python test.py -a Bangs,Smiling -t input_imgs/Bangs/0/01.jpg -s input_imgs/Bangs/1/02.jpg -r temp/temp.jpg -c 001

In addition, for the performance of the validation set during training, please refer to the sample_training and sample_training_multi_diff_x0 folders under model_trained/001_default, and refer to summaries for the value change of loss funciton.