Playing with Neural Style Transfer

Hello World! 😉

I’ve always liked Impressionist paintings: the visible brushstrokes, the light, and the way a scene can feel familiar and strange at once. So I tried making my own images in that spirit with neural style transfer.

The idea is to start with two images. The content image supplies the scene; the style image supplies visual patterns such as colour and texture. The generated image tries to keep the first while borrowing from the second.

I used a pretrained VGG-19 network to represent features from the images. Its earlier layers respond to simpler visual features, while deeper layers capture more complex ones. Optimising the generated image against those representations is what makes the mix possible.

For the experiment, I pulled content photos from my computer and paired them with paintings. The outputs kept recognisable parts of the original scenes while picking up colour and brushstroke patterns from the style images. Here are a few results:

There’s always room to tune the settings and let it run longer. For now, I like that the results sit somewhere between a photo, a painting, and a small machine-made surprise.