Hair Style and Color changing using improved version of StyleGAN2, deployed on interactive Graphical User Interface (GUI) with PyQt.

• Paper link at Korea Software Congress (KSC) 2023 is provided as https://dbpia.co.kr/pdf/pdfView.do?nodeId=NODE11705141.
  
• BibTeX is added.

Paper accepted to the Korea Software Congress (KSC) 2023, Oral Session

Clone repo and install requirements.txt in a Python>=3.7.0 environment.

git clone https://github.com/abbasmammadov/Artificial-Barber  # clone
cd Artificial-Barber
pip install -r requirements.txt  # install all the required dependencies

• Our model is built on improved version of StyleGAN2. However, since the original model is quite heavy, it has been very hard to bring the implementation down to a product level.
• Hence, we have made an improved preprocessing and deployed our model on a Graphical User Interface.
• If you are curious about our implementation, be sure to check our paper

Note: Please make sure that your device is equipped with NVIDIA GPU and CUDA CuDNN

To try our model with your own choice of images we provided 2 options

We recommend using this option for inference, especially for those who are unfamiliar with linux and bash commands.

As shown below, our GUI is easy to use and provides users variety of options interms of choosing target color and style, as well as inference speed.

Please upload an image containing only 1 person, otherwise the interface automatically rejects the image

• Run python main.py and upload your input image (from any directory). Please check whether your image fulfills the requirements shown below

• Then our GUI automatically runs the face.py code, and performs face-alignment.

• Check whether to use the "fast-inference" or the "Best Quality" option. Fast Inference runs lighter models to achieve results as fast as possible, compromising quality, whereas "Best Quality" ensures to yield best results, but takes relatively longer time to process.

• If you see a success message, move to selecting the color and style of your choice

  Options Provided To choose your target color

  • You can choose either simple colors from the options in the "basic" category, or combination of colors from our "combo" option

  • As an another option, we provided a color palette where you can select your desired customized color with our novel idea.

  Note: Please make sure to check the "Use custom color" option to use this option.

  • Moreover using the "Custom color image" you can upload an image of your choice and our model converts to the hair color of that image

  Options Provided to choose your target style

  • First, when you hover your mouse over the given styles, you will see an example image carrying that style
  • You can choose either simple styles or you can also upload an image of your choice, and our model converts to the hair style of that image

  Then all you need to do is, pressing the "Generate" button and see the magic!!

  After the result is done, you will see an output which looks like this.

  

• If you prefer using the terminal, please follow the steps given below

• Please make sure that the input image is placed in the same directory with face.py. Then execute

  python face.py

Then the aligned face will be stored in the directory img/input/. After that, all you need to do is executing the command below

  python inference.py --im_path1 {path_of_your_input_image} --im_path2 {path_of_the_image_with_your_target_style} --im_path3 {path_of_the_image_with_your_target_color} --output_dir {desired_path_to_store_output_images_and_their_masks}

Then the result output and its mask will be saved to your output path folder.

• Fast Inference with better parameter tuning and pre-processing
• Availability of Custom Color
• Deployment through Graphical User Interface

If you find our work interesting, please consider citing

Abbas Mammadov, Kaleb Mesfin Asfaw.(2023).Artificial Barber: Hair Color and Style transfer using GANs.한국정보과학회 학술발표논문집,(),487-489.

This code is based on Barbershop.

If you feel adventurous, check out the following official papers for a foundation on StyleGAN, and StyleGAN2.

• A style-based generator architecture for generative adversarial networks
• Analyzing and improving the image quality of StyleGAN.