Painting with gradients: the 2015 neural style paper, reimplemented

What it does

Feed it a subject photo and a style image; it outputs the photo re-rendered in the style of the painting. The implementation follows the original 2015 paper closely, using a pretrained VGG-19 network to separate and recombine content and texture.

The interesting bit

The author built two custom dependencies to make this work: DeepPy (a deep learning framework) and CUDArray (a CUDA-accelerated NumPy clone). That was the state of the art in 2015 — before PyTorch made this a ten-line script.

Key highlights

• Implements the original Gatys et al. “A Neural Algorithm of Artistic Style” paper verbatim
• Command-line interface: --subject and --style flags, plus tunable options via --help
• Requires VGG-19 weights from MatConvNet’s pretrained zoo
• Ships with a full example gallery including the Queen of Denmark in five artistic styles
• GPU acceleration via cuDNN through the CUDArray backend

Caveats

• Dependencies (DeepPy, CUDArray) are the author’s own projects and appear unmaintained; getting this running on modern CUDA versions may require archaeology
• No mention of memory requirements, processing time, or supported image sizes in the README

Verdict

Worth a look if you’re studying the original paper or maintaining legacy 2015-era deep learning code. For practical style transfer today, modern frameworks have made this far simpler.