lengstrom/fast-style-transfer

This project applies the visual style of a painting to a photograph, producing a new image that combines the content of your photo with the colors and brushwork of the artwork. For example, you can take a cityscape photo and make it look like it was painted in the style of a specific impressionist painting. The same technique works on video, processing each frame and reassembling the result. The approach works by training a neural network on a particular painting style. Once training is complete, applying that style to a new image takes about 100 milliseconds on a capable graphics card. Training takes 4 to 6 hours and requires a GPU. The project includes several pre-trained style models so you can use it without training from scratch. On the technical side, the library is written in Python and depends on TensorFlow, an open source machine learning framework. It also requires specific versions of Python libraries for image processing, and FFmpeg if you want to stylize video. Setup instructions cover both Windows and Linux using a tool called Anaconda to manage the Python environment and its dependencies. Three command-line scripts handle the main tasks: one for training a new style network, one for applying an existing style to images, and one for processing video. Each script accepts a set of parameters you pass from the terminal. The license restricts use to academic research. Commercial use requires separate permission from the author. If you use this in a paper, the README includes a citation block to reference the project.