deepdream
Google Research's original 2015 DeepDream notebook that turns any photo into a surreal, pattern-covered image by asking a neural network to amplify what it thinks it sees.
DeepDream is a small repository from Google Research that contains a single Jupyter notebook showing how to generate trippy, surreal images by running pictures through a neural network in an unusual way. It was published in 2015 alongside a blog post that explained the technique, which Google called Inceptionism.
The idea behind DeepDream is to take a trained image-recognition network and instead of using it to identify what is in a photo, feed it an image and ask it to amplify whatever patterns it thinks it sees. The network has learned to recognize features like eyes, fur, or architecture, so when it keeps reinforcing those patterns in an image, the result looks like the original photo covered in strange repeating shapes and faces. The more passes you run, the more extreme the visual distortion becomes.
The repository is minimal: it is essentially one notebook file with sample code and explanations. There is no installable package or command-line tool. You clone the repository, install the dependencies listed in the notebook, and run the cells to generate your own images. The code was written to work locally or to be read directly on GitHub.
This was an influential early demonstration of what neural networks had learned to see, and the images it produced went viral in 2015. The technique and the hashtag deepdream sparked a wave of community-generated artwork. The README is short and the project has not been developed further since its original release.
Where it fits
- Run the notebook on your own photos to generate surreal DeepDream-style images.
- Study the original technique to understand how trained neural networks visualize and amplify learned features.
- Use the code as a reference when learning how gradient ascent is applied to image generation.