PyTorch GAN · 2024
Pixzip
A generative image-compression model that extracts the most useful visual information and reduces image data by up to 93%.
Pixzip is an image-compression model built during MHacks 2024. A convolutional discriminator trains an autoencoder and ResNet-based generator to preserve the most important visual information in a compact latent representation.
The project was built in PyTorch with NumPy, OpenCV, and pandas. In a 24-hour build, it reduced image data by 93% and matched JPEG in some scenarios, earning the MHacks 2024 Optimization Track award.
The project also includes a showcase website built with Next.js, React, and Three.js. The next step is closing the remaining quality gap with JPEG and exposing Pixzip through a beta compression API.