zeldaMash

2017-04-30 · AI/ML
neural-networkzeldamusicaideepdreamrnnmachine-learninggenerative-art

zeldaMash

Video was generated by a neural network trained on images of fire - still needs a lot of work.

This project combines two neural network experiments to create an AI-generated audiovisual experience.

The Visual Component: DeepDream

The visuals were created using Google's DeepDream algorithm. DeepDream works by running an image classification neural network in reverse - instead of identifying what's in an image, it enhances patterns the network "sees," creating surreal, dream-like imagery.

For this project, the network was trained on images of fire, resulting in flame-like patterns appearing throughout the video.

Based on: Google DeepDream

The Audio Component: Recurrent Neural Network

The music was generated by a recurrent neural network (RNN) trained on 15 Zelda songs. RNNs are particularly good at learning sequential data like music, as they can remember context from earlier in a sequence.

The network learned patterns from classic Legend of Zelda compositions and generated new music that captures elements of the game series' iconic sound.

Based on: Biaxial RNN Music Composition

An Early AI Experiment

This was an early experiment in AI-generated multimedia content, exploring:

While the creator notes it "still needs a lot of work," this project represents the kind of experimentation that eventually led to today's more sophisticated AI art and music generation tools.

Watch on YouTube

View in Interactive Desktop App