Retratista is a web-based application that enables the user to generate and manipulate a realistic human face from bare description.
For further details about our experimentation, refer to our paper and experiments repo portrait-synthesis-from-speech.
System Workflow
Figure(1): Complete block diagram showing the whole system design and flow between different modules.
Application Design
Figure(2): Web application design.
Usage
Docker
-
Make sure you installed :
- Nvidia CUDA 11.1 + CUDNN8
- Docker CE.
- Docker Compose.
- Nvidia Docker V2
-
Start application server using docker-compose :
bash scripts/docker_run.sh base # to run stylegan2 server with ui bash scripts/docker_run.sh pose # to run pose server only bash scripts/docker_run.sh # to run both servers with ui
Native
-
Make sure you installed :
- Nvidia CUDA 11.1 + CUDNN8
- python3 + python3-pip (Anaconda can be used as well).
-
Install python dependencies :
pip3 install -r requirements.txt
-
Download model weights :
bash scripts/download_weights_gen.sh # to download weights for stylegan2 server bash scripts/download_weights_rot.sh # to download weights for pose server
-
Run application stylegan2 server (port 5000) :
python3 run_generator.py production 5000 # production mode python3 run_generator.py development 5000 # debug mode
-
Run application pose server (port 5001) :
python3 run_rotator.py production 5001 # production mode python3 run_rotator.py development 5001 # debug mode
-
Build and run application ui (port 8080) :
cd ui bash setup_ui.sh npm run style # only stylegan2 server # OR npm run styleRR # both stylegan2 and pose server