They do explain what they do: https://kundor.github.io/Cheating-images/
I find their methodology for evaluating candidate Waldos to be as sufficient as any other.
What I mean is, they could certainly choose a much harder evaluation criteria and fail, but that's not much of a blog.
In that case I'd be cool to know the probability in finding a close enough image in that optimized space
In the candidate stream, byte 0 is 4. The error is minimised if byte 2 is also 4. The values of the bytes don't matter, just the patterns of which bytes have the same value.
But they're not finding exact bytes, they're finding something that very vaguely looks like something very vague. It also helps a lot that it's a face and human brains are very good at identifying faces.
The number of 500 bytes that would "look like" waldo is a lot higher than 1.