Measuring the energy contribution of audio generation in production T2V systems

Ascertain the energy cost attributable to audio generation components within production text-to-video systems that also synthesize audio, and quantify their contribution relative to the video generation pipeline under realistic usage.

Background

The paper benchmarks energy and latency for video generation but does not include audio, even though many production systems generate synchronized audio alongside video.

The authors explicitly note that the energy contribution of audio remains unexplored, indicating the need to measure and compare audio’s share to GPU-dominated video inference to complete the end-to-end energy budget.