OTCache Accelerates Diffusion Models with Geometry-Aware Caching

A new framework called OTCache has been introduced to significantly speed up diffusion model sampling without requiring additional training. Current graph-based caching methods for diffusion models often struggle in low-NFE (Number of Function Evaluations) regimes due to an assumption of additive independence that doesn't always hold. OTCache addresses this by modeling caching schedules as a smooth evolution in policy space, drawing inspiration from Optimal Transport theory. The process involves three steps: first, establishing a high-fidelity reference schedule; second, conducting a lightweight anchor search for extreme low-budget settings; and finally, predicting schedules for target budgets through quantile interpolation. Experimental results demonstrate that OTCache achieves impressive acceleration rates, such as 4.5x on FLUX.1, 4.7x on Qwen-Image, and 3.66x on HunyuanVideo. Crucially, these speedups are achieved while consistently enhancing the quality of generated images compared to existing state-of-the-art caching methods. This work offers a novel perspective on optimizing diffusion models through advanced schedule modeling.