As suggested in the [Twitter thread](https://twitter.com/Sree_Harsha_N/status/1232747719581564928), let's try using [uber-research/PPLM](https://github.com/uber-research/PPLM) and compare the results