Neural Style TF is a TensorFlow implementation of neural style transfer, the technique that made headlines by repainting photographs in the style of famous paintings. Give it a content image and a style image, and its convolutional neural network separates and recombines the two, synthesizing a convincing pastiche.
The project is hosted on GitHub, where it has earned more than 3,100 stars under the GPL-3.0 license. It implements the methods from the foundational papers by Gatys, Ecker, and Bethge on image style transfer using convolutional neural networks, and extends them with later research: artistic style transfer for videos, and techniques for preserving color in neural artistic style transfer. Beyond the basics, the codebase presents semantic segmentation-guided styling, so different regions of an image can be treated intelligently, and support for transferring multiple styles. The repository famously reproduces the paper’s figure rendering a photograph of the Neckarfront in Tübingen in the styles of five iconic paintings, from The Starry Night to The Scream. The code was last updated in December 2020, so it reflects the TensorFlow era of the technique rather than today’s diffusion models, but that is precisely its value: it is a compact, readable reference implementation of the original algorithm that much modern creative tooling descends from, and it still runs for anyone who wants to understand or reproduce the classic results.
Neural Style TF is free and open source under the GPL-3.0 license. There are no fees; you run it on your own machine, ideally with a GPU for reasonable speeds.
It is for students, computer vision researchers, and creative coders who want to learn how neural style transfer actually works from a faithful implementation, or to reproduce the classic paper results, rather than calling a modern black-box image API.