Open-source team compresses 31B video-audio diffusion model to 14B — faster, cheaper generation

Context

A developer project posted to Hacker News claims to have compressed a 31-billion-parameter video and audio diffusion model down to 14 billion parameters, running in six sampling steps with INT8 quantisation. The release is a community Show HN post with modest traction, so the claims have not been independently verified. If the performance holds, lighter generative video models could lower the cost of producing AI video assets for brand content, which is worth watching for creative and production teams.