StreamDiffusion at event speed
The technical constraint was latency. Most AI image generation is built around batch processing, which introduces delays that break the illusion when you are standing in front of a mirror. StreamDiffusion runs continuous frame-by-frame diffusion, which let us bring the update rate close enough to real time that visitors saw themselves transforming as they moved.We ran the full pipeline on on-site GPU hardware, no cloud dependency, because a dropped connection at a live event with a queue waiting is not acceptable. The system ran across both days of the festival without a failure.
