>>84064303I am that anon and it looks like this timestep schedule isn't even the best one.
>>>/h/8175944SpeeD's base timestep schedule was pretty straightforward to implement if you yank their implementation directly, but Change-Aware Weighting was kind of a pain honestly.