About the learning method of neural_style_transfer

I don't understand the content of program of neural_style_transfer, especially the parameters of optimizer. Why the params of optimizer is image, nor model's paramters? and how this program get such good results? what's the learning method of it? if you can, please give me the corresponding article. Thanks very much!(*^_^*)