Artistic work is generally recognized through an artist’s brush strokes, color choices, color palette and so on. Each artist’s style is unique and hard to replicate. With advancements in computer vision however, Generative Adversarial Networks are now able to mimic such nuances in a very convincing way. We intend to use GANs (specifically Cycle GAN) to replicate Claude Monet’s artistic style and see if our GAN can create realistic renditions of his work by transforming normal photos, as well as paintings by other famous artists.
You can run the base code in two ways:
- Run train.py in cyclegan and then test.py
- Run the colab notebook cgan_base_code.ipynb
test.py has code that can pull in saved weights.
Other colab notebooks can be run too to experiment with different artists and datasets.
To run the model weights to convert a normal video to monet style video:
cd video-lib
python main.py video path-to-videoFor webcam:
cd video-lib
python main.py webcamNOTE: works best on nature/scenery video frames
For reference, there's already a generated video in video-lib: