MeanCache: From Instantaneous to Average Velocity for Accelerating Flow Matching Inference
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866911497883484160 |
|---|---|
| author | Gao, Huanlin Chen, Ping Shi, Fuyuan Wu, Ruijia YanTao, Li Hui, Qiang You, Yuren Lu, Ting Tan, Chao Zhao, Shaoan Liu, Zhaoxiang Zhao, Fang Wang, Kai Lian, Shiguo |
| author_facet | Gao, Huanlin Chen, Ping Shi, Fuyuan Wu, Ruijia YanTao, Li Hui, Qiang You, Yuren Lu, Ting Tan, Chao Zhao, Shaoan Liu, Zhaoxiang Zhao, Fang Wang, Kai Lian, Shiguo |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2601_19961 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | MeanCache: From Instantaneous to Average Velocity for Accelerating Flow Matching Inference Gao, Huanlin Chen, Ping Shi, Fuyuan Wu, Ruijia YanTao, Li Hui, Qiang You, Yuren Lu, Ting Tan, Chao Zhao, Shaoan Liu, Zhaoxiang Zhao, Fang Wang, Kai Lian, Shiguo Machine Learning Artificial Intelligence Computer Vision and Pattern Recognition 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. |
| title | MeanCache: From Instantaneous to Average Velocity for Accelerating Flow Matching Inference |
| topic | Machine Learning Artificial Intelligence Computer Vision and Pattern Recognition |
| url | https://arxiv.org/abs/2601.19961 |