Trajectory-Consistent Calibration for Cache-Accelerated Diffusion Models 事件

Trajectory-Consistent Calibration for Cache-Accelerated Diffusion Models arXiv:2605.24870v1 Announce Type: new Abstract: Diffusion Transformers require repeated denoiser evaluations during iterative sampling, making inference computationally expensive. Cache-based acceleration reduces this cost by reusing intermediate representations across denoising steps, but can introduce representation deviations and degrade generation quality. In this paper, we analyze these deviations and show that effect