MeanCache: From Instantaneous to Average Velocity for Accelerating Flow Matching Inference
This addresses efficiency and quality issues in generative models like FLUX.1, Qwen-Image, and HunyuanVideo, offering a novel perspective for Flow Matching inference, though it is incremental as it builds on existing caching methods.
The paper tackles the problem of error accumulation in caching methods for Flow Matching inference by introducing MeanCache, a training-free framework that uses average velocities from cached Jacobian-vector products, achieving accelerations of 4.12X, 4.56X, and 3.59X on three datasets while outperforming state-of-the-art baselines in generation quality.
We present MeanCache, a training-free caching framework for efficient Flow Matching inference. Existing caching methods reduce redundant computation but typically rely on instantaneous velocity information (e.g., feature caching), which often leads to severe trajectory deviations and error accumulation under high acceleration ratios. MeanCache introduces an average-velocity perspective: by leveraging cached Jacobian--vector products (JVP) to construct interval average velocities from instantaneous velocities, it effectively mitigates local error accumulation. To further improve cache timing and JVP reuse stability, we develop a trajectory-stability scheduling strategy as a practical tool, employing a Peak-Suppressed Shortest Path under budget constraints to determine the schedule. Experiments on FLUX.1, Qwen-Image, and HunyuanVideo demonstrate that MeanCache achieves 4.12X and 4.56X and 3.59X acceleration, respectively, while consistently outperforming state-of-the-art caching baselines in generation quality. We believe this simple yet effective approach provides a new perspective for Flow Matching inference and will inspire further exploration of stability-driven acceleration in commercial-scale generative models.