lengstrom/fast-style-transfer ? reverse-engineered prompt

Build me a TensorFlow project that can take a normal photo and make it look like a famous painting really fast, and also do the same for videos. I want a training script that learns a style from one image, an evaluation script that can stylize a folder of photos using a saved model, and a video script that can process an input video and write out a stylized version. Please include simple setup instructions, a way to download or use sample content and style images, and a few example commands so I can try it right away. It should work with GPU if available, but still run on CPU if needed, just slower. Also add clear notes for where to put checkpoints and how to run the scripts from the command line. If you need to, look up any current TensorFlow or ffmpeg docs online so the instructions are easy to follow.