Paper

Beyond Fixed Formulas: Data-Driven Linear Predictor for Efficient Diffusion Models

Cai, Peiliang, Shen, Zhirong, Zou, Chang, Zheng, Shikang, Liu, Jiacheng, Liang, Feng, Huang, Rui, Zhang, Linfeng, Shi, Zhengyi

2026

Overview

To address the high sampling cost of Diffusion Transformers (DiTs), feature caching offers a training-free acceleration method. However, existing methods rely on hand-crafted forecasting formulas that fail under aggressive skipping. We propose L2P (Learnable Linear Predictor), a simple data-driven caching framework that replaces fixed coefficients with learnable per-timestep weights. Rapidly trained in ~20 seconds on a single GPU, L2P accurately reconstructs current features from past trajectories. L2P significantly outperforms existing baselines: it achieves a 4.55x FLOPs reduction and 4.15x latency speedup on FLUX.1-dev, and maintains high visual fidelity under up to 7.18x acceleration on Qwen-Image models, where prior methods show noticeable quality degradation. Our results show learning linear predictors is highly effective for efficient DiT inference. Code is available at https://github.com/Aredstone/L2P-Cache.

• Publisher: Cornell University Library, arXiv.org
• Subject: Caching