New Method Accelerates Diffusion Model Generation Without Retraining.

Diffusion and flow matching models are excellent at generating high-quality samples, but their reliance on ODE samplers often requires many neural function evaluations (NFEs), making them slow. Existing acceleration techniques typically demand retraining or distillation, which adds complexity and cost. This research proposes a novel, training-free approach called Truncated Jump Sampling (TJS). TJS exploits a property called "endpoint decodability," where intermediate states in the sampling process already contain information about the final clean sample. By formalizing this as the minimum-MSE estimator, TJS allows the sampling process to be stopped early at an optimal time, and the clean sample is then decoded. This method requires no changes to the model architecture or additional training. Evaluations across models like SDXL, SD3.5M, and Z-Image-Turbo show that TJS can reduce NFEs by 20-70% while maintaining nearly identical sample quality. This offers a practical solution for accelerating existing diffusion model checkpoints without the overhead of retraining or trajectory redesign.