wtf is lossy?

This project takes a whole movie and compresses it into a small text file, then uses AI video and audio generation to rebuild a new version of the film from that text. It calls itself a lossy codec, meaning the rebuilt film is not an exact copy, it is a reconstruction based on written descriptions of what happens in each shot. The process works in stages. First the film is split into roughly two thousand shots using scene detection tools, and keyframes are pulled from each one. Next, an AI vision model looks at sampled frames and writes a detailed description of each shot, including things like camera movement and sound. Those descriptions, plus some audio classification, get bundled into a single JSON file that ends up under one megabyte once compressed. That tiny file is effectively the whole film boiled down to text. To turn the text back into video, the project sends each shot's description to an AI video generation model, which can be run on a rented cloud GPU or through a paid API service, and several different backend options are supported with different costs and speed tradeoffs. Sound effects and music are generated separately by other AI audio models, and any spoken dialogue is recreated using text to speech based on the film's original subtitles. Finally all the generated clips and audio are stitched back together into one output video, and a browser based tool lets you compare the original and reconstructed versions side by side. According to the author's write-up, reconstructing a full feature film this way costs around thirty dollars in rented GPU time plus about a dollar for the description step. The project is written in Python, needs FFmpeg and the uv tool to install its dependencies, and requires API keys for the AI services it depends on. It is released under the MIT license.