Easy there. You've misunderstood the benchmark. It shows the time to add 100 images, given the existing number of images. So adding the 900th-1000th image was ~3x faster than adding the 100th-200th.
So if the benchmark produces weird results, and is published without source code, so that the results cannot be reproduced, why would anyone trust it?
Even the author suspects the V8 optimized away the vanilla code... and if that is what happened, then apples are being compared to oranges and the whole conclusion is bogus. Which was kind of my point.