E²-CRF method accelerates frequency domain diffusion models for time series generation with ~2.2 speedup

Researchers propose E²-CRF (Error-Feedback Event-Driven Cumulative Residual Feature caching), a technique that speeds up diffusion models (AI systems that generate time series data through iterative refinement) by exploiting spectral localization and mirror symmetry in frequency domain signals.

The method uses a closed-loop error-feedback system to adaptively cache transformer KV features across diffusion steps, triggering recomputation only when needed via event-driven residual dynamics rather than fixed schedules, selectively recomputing high-energy or rapidly-changing tokens while reusing cached features for stable components.

E²-CRF achieves ~2.2 speedup while maintaining sample quality and was validated on 5 datasets; code is available online and integrated into an existing repository.