跳到正文
返回论文列表
cs.CV提交于 已译

CalibCache:面向扩散 Transformer 的缓存到全量特征动态校准

CalibCache: Cache-to-Full Feature Dynamics Calibration for Diffusion Transformers

Mingyu Liang · Dingkun Xu · Jingwei Xu

中文摘要

扩散 Transformer(Diffusion Transformer)能生成高质量图像,但迭代采样计算开销大。特征缓存可减少全网络评估次数,但在短采样调度下近似误差更严重:全计算步更少、采样网格更粗,使基于历史的外推更不可靠。CalibCache 是一种轻量离线校准方法,用于修正缓存与全计算特征动态之间的系统性偏差。它校准特征加速度,即跨采样步的二阶有限差分。每个设置下仅用一对全计算与朴素缓存轨迹,通过中心最小二乘离线拟合标量加速度映射;采样时直接应用,无需额外全网络评估或在线优化。在 DrawBench 上的 FLUX.1-dev 与 Qwen-Image、ImageNet-256 上的 DiT-XL/2 上,20 步采样、缓存间隔 N=3,4,5 时,CalibCache 在保真度或生成质量上稳定优于 TaylorSeer。N=5 时,FLUX.1-dev 的 LPIPS 从 0.563 降至 0.429,去噪相对全计算采样加速 2.86 倍;DiT-XL/2 的 FID 相对 TaylorSeer 从 16.39 降至 10.37。

关键要点

  1. 01短采样调度下缓存近似误差更严重,历史外推可靠性下降
  2. 02离线拟合标量加速度映射,校准缓存与全计算特征动态偏差
  3. 03采样时直接应用校准,无需额外全网络评估或在线优化
  4. 04N=5 时 FLUX.1-dev LPIPS 从 0.563 降至 0.429,加速 2.86 倍
  5. 05DiT-XL/2 FID 相对 TaylorSeer 从 16.39 降至 10.37

解读

尚无解读。

原始英文摘要

arXiv:2611.25296v1 Announce Type: new Abstract: Diffusion Transformers generate high-quality images, but iterative sampling remains computationally expensive. Feature caching reduces the number of full network evaluations, yet its approximation error can become more severe under short sampling schedules, where fewer full-compute steps and a coarser sampling grid can make history-based extrapolation less reliable. In this paper, we propose CalibCache, a lightweight offline calibration method that corrects systematic discrepancies between cached and full-compute feature dynamics. Specifically, CalibCache calibrates feature acceleration, defined as the second-order finite difference across sampling steps. Using only a single pair of full-compute and naive-cache trajectories at each setting, CalibCache fits scalar acceleration maps offline by centered least squares. These maps are then applied during sampling without additional full-network evaluations or online optimization. Across FLUX.1-dev and Qwen-Image on DrawBench and DiT-XL/2 on ImageNet-256, CalibCache consistently improves fidelity or generation quality over TaylorSeer under 20-step sampling at cache intervals $N=3,4,5$. At $N=5$, it reduces FLUX.1-dev LPIPS from $0.563$ to $0.429$ while achieving a $2.86\times$ denoising speedup over full-compute sampling, and lowers DiT-XL/2 FID from $16.39$ to $10.37$ relative to TaylorSeer.

同方向论文 · cs.CV

查看全部 →