Training-free adaptive diffusion with bounded difference approximation strategy