Recently, it was reported that gaussian distribution and difference of two gaussians (DOG) to have band pass characteristics can be generated by simple iterative processes of the diffusion networks. In this paper, we propose method of improved implementation of a diffusion filter which can reduce total runing time, and operate by simple algorithm in contrast to the latest diffusion filter. We rebuild the diffusion network to a difference network which can generate DOG independently. Different filter characteristics are obtained just by each diffusion process and difference process. Quantitative analysis shows that the center frequency and the selectivity of each filter channel can be varied independently. Also, it would requires smaller amount of hardwares than conventioanl method to build a filter bank.