Deep Hierarchical Planning from Pixels

In the results graphs it looks like it gets beat pretty bad by the “dreamer” baseline, am I missing something?