超越固定公式：面向高效扩散模型的数据驱动线性预测器
Beyond Fixed Formulas: Data-Driven Linear Predictor for Efficient Diffusion Models
把手工设计的 Taylor 式特征外推系数换成按时间步学习的线性权重，只需少量缓存轨迹、单卡约 20 秒即可训好，可直接插到 DiT/FLUX/Qwen-Image 等现有采样管线上做免训练加速，是特征缓存类加速方法的一个即插即用替换件。
原文依据 we propose L^2P (Learnable Linear Predictor), a simple data-driven caching framework that replaces hand-designed coefficients with learnable per-timestep weights
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论扩散 Transformer（DiT）在图像与视频生成上取得了最优性能，但由于需要在众多时间步上反复进行 Transformer 前向计算，采样依然开销高昂。特征缓存提供了一种免训练的推理加速途径，通过复用或预测隐藏表征来减少计算，然而近期基于预测的方法其系数来自手工设计的公式（如 Taylor 展开），最终都归结为少数历史特征的固定线性组合。这类固定系数是次优的，并且在激进跳步时十分脆弱。本文首先证明现有基于预测的缓存方法可以统一为一种共同的线性形式，随后分析 DiT 的特征轨迹，发现在大多数去噪步上，当前特征可由历史特征重建，投影保真度高于 0.95，这表明精确的线性预测是可行的。受此启发，我们提出 L^2P（可学习线性预测器），一个简单的数据驱动缓存框架，它用在一小批缓存轨迹上以均方误差损失训练得到的、按时间步可学习的权重，取代手工设计的系数，在单块 GPU 上约 20 秒即可收敛。在最优 DiT 上的大量实验表明，L^2P 持续优于现有缓存基线：在 FLUX.1-dev 上，L^2P 实现 4.55 倍 FLOPs 降低与 4.15 倍延迟加速，峰值信噪比为 31.459；在 Qwen-Image 与 Qwen-Image-Lightning 上，即便加速高达 7.18 倍，它仍能保持较高的视觉保真度，而此前的方法则出现明显的质量退化。这些结果表明，学习线性预测器是设计日益复杂的预测公式的一种实用且有效的替代方案，可用于高效的扩散模型推理。实现代码已发布于 https://github.com/Aredstone/L2P-Cache。
英文原文摘要 1916 字符
Diffusion Transformers (DiTs) have achieved state-of-the-art image and video generation performance, but sampling remains expensive due to repeated transformer forward passes over many timesteps. Feature caching offers a training-free way to accelerate inference by reusing or forecasting hidden representations, yet recent forecasting-based methods derive their coefficients from hand-crafted formulas (e.g., Taylor expansion), which ultimately reduce to fixed linear combinations of a few historical features. Such fixed coefficients are suboptimal and fragile under aggressive skipping. In this paper, we first show that existing forecasting-based caching methods can be unified in a common linear form, and then analyze DiT feature trajectories, finding that for most denoising steps the current feature can be reconstructed from past features with projection fidelity above 0.95, indicating that accurate linear prediction is feasible. Motivated by this, we propose L^2P (Learnable Linear Predictor), a simple data-driven caching framework that replaces hand-designed coefficients with learnable per-timestep weights trained on a small set of cached trajectories using a mean-squared error loss, converging in about 20 seconds on a single GPU. Extensive experiments on state-of-the-art DiTs demonstrate that L^2P consistently outperforms existing caching baselines: on FLUX.1-dev, L^2P achieves a 4.55xFLOPs reduction and 4.15xlatency speedup with a PSNR of 31.459, and on Qwen-Image and Qwen-Image-Lightning, it maintains high visual fidelity even under up to 7.18xacceleration, where prior methods suffer from noticeable quality degradation. These results show that learning linear predictors is a practical and effective alternative to designing increasingly complex forecasting formulas for efficient diffusion model inference. The implementation code is available at https://github.com/Aredstone/L2P-Cache.