用可学习权重替代固定公式，让扩散模型采样快4倍以上
Beyond Fixed Formulas: Data-Driven Linear Predictor for Efficient Diffusion Models
- 用数据驱动方法替换手工公式，每步预测用可学权重
- 在FLUX.1-dev上减少4.55倍计算量，提速4.15倍
- 支持7.18倍加速仍保持画质，适合高效率图像生成
为降低扩散变换器（DiTs）的采样开销，特征缓存提供了一种无需训练的加速方法。但现有方法依赖手工设计的预测公式，在激进跳步下表现不佳。本文提出L2P（可学习线性预测器），一种数据驱动的缓存框架，将固定系数替换为每步可学习的权重。仅需单卡20秒即可快速训练，能准确从历史轨迹重建当前特征。L2P显著优于现有基线：在FLUX.1-dev上实现4.55倍FLOPs减少和4.15倍延迟加速；在Qwen-Image模型上，即使加速达7.18倍，仍保持高视觉保真度，而此前方法出现明显质量下降。结果表明，学习线性预测器对高效DiT推理极为有效。代码已开源：https://github.com/Aredstone/L2P-Cache。
原文摘要 · Abstract (English)
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.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。