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

TL;DR

This paper introduces L2P, a data-driven linear predictor that significantly accelerates diffusion models by replacing fixed formulas with learnable weights, achieving substantial speedups with maintained quality.

Contribution

L2P is a simple, fast-to-train, learnable linear predictor framework that improves diffusion model inference efficiency over fixed formula methods.

Findings

L2P achieves 4.55x FLOPs reduction on FLUX.1-dev.

L2P maintains high visual fidelity under 7.18x acceleration on Qwen-Image.

L2P outperforms existing caching baselines in speed and quality.

Abstract

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.