SoundWeaver: Compositional Warm-Starting for Text-to-Audio Diffusion Serving

Ayush Barik, Sofia Stoica, Nikhil Sarda +4

Text-to-audio (T2A) diffusion models generate high-quality audio but require tens of neural function evaluations (NFEs), resulting in substantial inference latency. Existing accele…