Abstract

Diffusion models have achieved remarkable generative quality but remain bottlenecked by costly iterative sampling. Recent training-free methods accelerate diffusion process by approximating denoising trajectories. However, these methods primarily control the speedup through preset approximation frequencies, lacking an online assessment of how approximation affects the generation result. We conduct an empirical study and find that diffusion models exhibit a self-correcting behavior, where model outputs compensate for accumulated deviations. Leveraging this observation, we introduce Estimate Then Control (ETC). Rather than estimating trajectories from model outputs entangled with error corrections, ETC performs extrapolation in predicted-x0 space, where accumulated deviations are compensated during sampling and enable more reliable cross-step prediction. To further assess the impact of the current approximation strategy, we design a risk indicator that jointly measures the relative correction magnitude and directional consistency of predicted-x0 evolution. The approximation window is then dynamically adapted to prevent accumulated errors from exceeding the model's self-correction capacity. Experiments show that ETC achieves a 3.2× acceleration over FLUX with negligible (-0.094 SSIM score) degradation of consistency.

 

Method

Overview of ETC. Subfigure (a) depicts the acceleration process, which dynamically skips model predictions. Subfigure (b) illustrates how skipped denoising steps are approximated by extrapolating the self-corrected predicted-x0 trajectory from the recent evolution. Subfigure (c) shows how the next approximation window is adapted based on the accumulated error.

 

Results