Training a Quantized Video Compression Autoencoder

Follow the full discussion on Reddit.
The aim of this project is to achieve extreme video compression using a quantized autoencoder combined with other generative models that can enhance the quality of the decompressed video. The motivation for this project is the need for a fast and efficient compression algorithm that can handle large amounts of video data from the web, which current hard-coded algorithms cannot satisfy (my needs, as I store a LOT of data). Moreover, this project explores the potential of this model architecture for long video generation, by integrating a CLIP-like structure to condition the frames.