The obvious problem is the relative scale of the pixelation to the underlying redacted material. The demo works because the scale encodes enough data of each character and its relationship to its neighbours. My guess would be even a 25% larger pixelation block would scrub too much for it to be at all reliable.
Then you move to guessing by syllables or whole words. Probably an order of magnitude or larger brute force job that way, but it should be possible. Especially if you target common words in the sentence first.
Photos won't work with this, it works here because text is discrete and provides a limited search space.
“Attacks always get better, they never get worse“ —Bruce Schneier
Yes, but that doesn't mean this particular attack is relavant to the problem space.
What makes it possible to search over font characters but not, eg, convolution tiles?
Photo identification is done with a limited search space of “tiles” that the image is decomposed into, for convolutional NNs.