ResilPhase Accelerates Diffusion Models with Noise-Resilient Extrapolation.

Diffusion models, while powerful, are often slow due to high inference latency. Existing acceleration methods, like "cache-then-forecast" schemes, struggle with quality degradation at aggressive speeds because they extrapolate discrete representations that are misaligned and numerically unstable. This leads to accumulated errors and amplified noise. ResilPhase addresses these challenges by reframing accelerated inference as stable macro-trajectory extrapolation within the ordinary differential equation (ODE) space. Instead of predicting intermediate features, it aligns forecasting with the model's Global Drift, ensuring consistency and reducing memory overhead. To combat the inherent noisiness of higher-order temporal derivatives, ResilPhase employs a derivative-free barycentric Lagrange extrapolator. Additionally, a bounded Phase Mapping regularizes the extrapolation domain, preventing oscillatory error growth. This combined approach allows ResilPhase to achieve state-of-the-art fidelity even under aggressive acceleration ratios, as demonstrated on models like FLUX.1-dev and HunyuanVideo.