✨ Accepted by ICLR 2026 ✨
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 aver age velocities from instantaneous velocities, it effectively mitigates local error accumulation. To further improve cache timing and JVP reuse stability, we de velop a trajectory-stability scheduling strategy as a practical tool, employing a Peak-Suppressed Shortest Path under budget constraints to determine the sched ule. Experiments on FLUX.1, Qwen-Image, and HunyuanVideo demonstrate that MeanCache achieves 4.12×, 4.56×, and 3.59× acceleration, respectively, while consistently outperforming state-of-the-art caching baselines in generation qual ity. 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.
In Flow Matching inference, existing caching methods primarily rely on reusing Instantaneous Velocity or its feature-level proxies. However, we observe that instantaneous velocity often exhibits sharp fluctuations across timesteps. This leads to severe trajectory deviations and cumulative errors, especially as the cache interval increases.
Inspired by MeanFlow, we propose MeanCache. Compared to unstable instantaneous velocity, Average Velocity is significantly smoother and more robust over time. By shifting the caching perspective from a single "point" to an "interval," MeanCache effectively mitigates trajectory drift under high acceleration ratios.
Figure 1: Instantaneous vs. Average Velocity and JVP Caching. (Left) Along the original trajectory, instantaneous velocity shows sharp fluctuations, while average velocity is much smoother. (Middle) At timestep 927, JVP Caching reduces error accumulation, though its effectiveness depends on the cache interval and hyperparameter $K$. (Right) At timestep 551, it achieves stronger error mitigation, showing that effectiveness varies across timesteps. Both middle and right figures are under the single-cache setting on the original trajectory.
We leverage Jacobian-Vector Products (JVP) to construct estimated interval average velocities. By reusing JVP information calculated at prior timesteps, we transform the current instantaneous velocity $v$ into an estimated average velocity $\hat{u}$ for the target interval. This approach compensates for trajectory deviations without additional inference overhead:
Figure 2: From Instantaneous to Average Velocity. Directly caching the instantaneous velocity $v(z_t,t)$ over $[t,s]$ easily leads to trajectory drift and error accumulation, whereas the average velocity $u(z_t,t,s)$ accurately reaches the target $s$. MeanCache introduces a prior timestep $r$ and reuses $\mathrm{JVP}_{r \to t}$ to estimate the average velocity $\hat{u}(z_t,t,s)$, thereby correcting the trajectory and effectively mitigating error accumulation.
To achieve an optimal balance between speed and quality while accounting for temporal heterogeneity, we propose a trajectory-stability scheduling algorithm:
| Method | Z-Image-base | MeanCache (B=25) | MeanCache (B=20) | MeanCache (B=15) | MeanCache (B=13) |
|---|---|---|---|---|---|
| Latency | 18.07 s | 9.15 s | 7.36 s | 5.58 s | 4.85 s |
| T2I |
|
|
|
|
|
Comparisons on a Single H800 GPU
Maintaining content consistency is a primary challenge for acceleration frameworks. Rare words, characterized by ambiguous semantics and low frequency, often lead to significant visual drift during the denoising process. MeanCache demonstrates superior potential in addressing this challenge.
Figure 3: Content consistency under rare-word prompts "Matutinal".
This benchmark evaluates the performance of MeanCache against LeMiCa using the Qwen-Image-2512 model as the base.
Baseline Latency (Original Qwen-Image-2512): 32.8s
| Constraint | Method | Latency | Speedup | Time Reduction |
|---|---|---|---|---|
| $B=25$ | LeMiCa | 18.83 s | 1.74x | - |
| MeanCache | 17.13 s | 1.91x | 9.0% | |
| $B=17$ | LeMiCa | 14.35 s | 2.29x | - |
| MeanCache | 11.67 s | 2.81x | 18.7% | |
| $B=10$ | LeMiCa | 10.41 s | 3.15x | - |
| MeanCache | 6.95 s | 4.72x | 33.2% |
| Constraint | Method | PSNR (↑) | SSIM (↑) | LPIPS (↓) |
|---|---|---|---|---|
| $B=25$ | LeMiCa | 29.20 | 0.945 | 0.065 |
| MeanCache | 29.46 | 0.944 | 0.057 | |
| $B=17$ | LeMiCa | 24.31 | 0.835 | 0.176 |
| MeanCache | 26.49 | 0.907 | 0.104 | |
| $B=10$ | LeMiCa | 17.80 | 0.637 | 0.368 |
| MeanCache | 19.44 | 0.767 | 0.237 |
@inproceedings{gao2025meancache,
title = {MeanCache: From Instantaneous to Average Velocity for Accelerating Flow Matching Inference},
author = {Huanlin Gao and Ping Chen and Fuyuan Shi and Ruijia Wu and Yantao Li and Qiang Hui and Yuren You and Ting Lu and Chao Tan and Shaoan Zhao and Zhaoxiang Liu and Fang Zhao and Kai Wang and Shiguo Lian},
booktitle = {International Conference on Learning Representations (ICLR)},
year = {2026},
url = {https://arxiv.org/abs/2601.19961}
}