There are many papers try to build models to generate images. Some of them try to tackle the problem from different point of views (probability or energy). Trainings fail easily if bad parameters are used. I want to talk about how the parameters (e.g. large learning rate) affect the trainings.