EchoCache: Energy-Guided Cross-Modal Caching for Efficient Audio-Driven Video Generation

Researchers introduced EchoCache, an energy-guided cross-modal caching framework for audio-driven video generation.

EchoCache uses audio time-frequency energy to decide when to update latent caches, plus dynamic timestep-latent caching and quantized cache management.

The method targets the high inference cost of diffusion-based A2V generation while preserving quality and audio-visual alignment.

It reports up to a 2.46× speedup on a mainstream model, improving the latency-quality trade-off for practical use.