Reproduce the bad result with vanilla diffusion objective

Hi! Thank you for the excellent work. This work identifies and provides a solution to a problem that has troubled the community for a long time. 

Could you please provide a configuration/script to reproduce the bad training result w/ diffusion objective shown in the paper?

