A Python-based live webcam demo of real time artistic style transfer using neural networks. The user can select between several different artistic styles and see the style applied to his or her live webcam video stream in near-realtime. Styled images from the live stream image can also be captured and shared.
A kiosk demo has been on display at Booz Allen Hamilton's Innovation Center.
Screenshot in windowed mode.
Perform a recursive clone option to automatically pull the chainer-fast-neuralstyle submodule.
git clone --recursive https://github.com/booz-allen-hamilton/neural-style-mirror.git
python nsmirror.py
Settings can be configured by editing settings.py.
Several style models trained specifically for this demo can be found in the models folder.
Any model compatible with chainer-fast-neuralstyle can be added to the models folder. A collection of additional style models trained by others can be found at: chainer-fast-neuralstyle-models.
This has been tested successfully with Python 2.7 on Linux and Windows.
- See
requirements.txtfor requirements that can be installed using pip/conda. - A webcam compatible with the installed version of OpenCV.
- To enable GPU mode, CUDA and CUDNN must be properly installed. Set GPU to the desired GPU in
settings.py. Without enabling GPU mode, this can be VERY slow.
The artistic styling methodology using feed-forward convolutional neural networks was proposed in "Perceptual Losses for Real-Time Style Transfer and Super-Resolution" by Justin Johnson, Alexandre Alahi, and Li Fei-Fei.
This demo utilizes Yusuke Tomoto's chainer-fast-neuralstyle implementation of the artistic style transfer algorithm.
Initial inspiration drawn from Gene Kogan's CubistMirror installation on display at the alt-AI conference.