GeoSPRINT Accelerates Diffusion Model Inference by Pruning Steps.

A new training-free framework called GeoSPRINT (Geometric Step Pruning for Inference in Trajectories) has been introduced to accelerate the inference process in diffusion models. Diffusion models are known for generating high-quality samples but often require numerous sequential neural function evaluations (NFEs), making inference computationally expensive. Existing acceleration methods typically involve fixed step-skipping schedules, adaptive step sizes based on local error, or additional training. GeoSPRINT, however, leverages the geometry of denoising trajectories to create more efficient, non-uniform sampling schedules. The core of GeoSPRINT lies in its ability to detect geometrically redundant steps within the latent space using a hyperplanarity test, efficiently implemented via QR factorization. This analysis generates a redundancy profile, which is then converted into a sampling schedule that allocates more steps to regions of high trajectory curvature and fewer to redundant, straighter segments. The framework also introduces a trajectory projection score, α_traj, a model-free metric for quantifying trajectory straightness and assessing rectified flow quality. Across various benchmarks, including CIFAR-10, LSUN Church, and Stable Diffusion v1.5, GeoSPRINT consistently outperformed uniform DDIM schedules at equivalent NFE budgets, showing significant improvements in FID scores and even surpassing DPM-Solver++ in some cases without requiring any retraining.