CacheScope: Exact-Compute Allocation for Training-Free Diffusion Acceleration
CacheScope is a training-free scheduling rule that optimizes where to perform full denoising computation during diffusion transformer inference. It addresses the problem of selecting high-risk steps by normalizing probe variation against input movement, creating a directional sensitivity score. This allows the system to selectively use exact computation only when it provides the most significant fidelity gain, leading to up to 3.57x acceleration.