Tag
This paper introduces Truncated Jump Sampling (TJS), a training-free method that accelerates diffusion and flow matching model generation by exploiting endpoint decodability, reducing neural function evaluations by 20–70% with near-matched quality across multiple models.