Neural Style Transfer

Neural Style Transfer is my implementation of the seminal paper "A Neural Algorithm of Artistic Style" by Gatys et al. The core idea is deceptively elegant: separate the content of an image from its style, then recombine them. A photo of a Berkeley street rendered in the brushstrokes of Van Gogh.

How it works

The algorithm uses a pretrained VGG-19 network as a feature extractor. Content is captured by the deeper convolutional layers, which respond to high-level structure. Style is captured by computing Gram matrices across multiple layers — these measure correlations between feature maps and encode texture information without caring about spatial position.

Optimization runs on the output image itself, minimizing a weighted combination of content loss and style loss via gradient descent. The balance between the two determines whether the result leans photorealistic or painterly.

What I built

• Full implementation in PyTorch with configurable content/style weights
• Support for arbitrary resolution inputs with GPU acceleration
• Experimented with different VGG layer combinations to see how feature depth affects style capture
• Added a total variation loss term to reduce high-frequency noise in outputs

The most interesting part was tuning the layer weights — shallower style layers capture fine textures like brushstrokes, while deeper ones transfer broader color palettes and compositional patterns.