jcjohnson/neural-style

Neural-style is an implementation of a technique that combines the visual content of one photograph with the artistic style of another image using neural networks (a type of AI modeled loosely on the brain). The underlying idea, from a research paper by Leon Gatys, Alexander Ecker, and Matthias Bethge, is that a neural network can separately identify what is in an image (its content) and how it looks (its style), then blend content from one source with style from another. The result is a new image that looks like the content photo repainted in the manner of the style image. For example, you could take a photograph of a city street and render it as if painted in the brushstroke style of a famous painting. The user can control how much weight to give the content versus the style, producing results that range from a barely-stylized photo to something almost entirely abstract. You can also blend multiple style images together, controlling each one's contribution, or keep the original colors of the content image while adopting only the texture and pattern of the style. Built using the Torch numerical computing framework and written in the Lua programming language, this was an early and influential implementation that demonstrated the technique's potential. Running it requires a graphics card for practical speed, and various pretrained neural network models for feature extraction.