I've recently read about NeuralHash, and immediately thought that it might be used as a loss for an autoencoder. However, it only seems to preserve "structure" from what I've read, not actual pixel values (which makes sense, given its purpose). Thus, how likely it is that an autoencoder performs well given a loss that compares the NeuralHash of its output with the NeuralHash of its input?
I feel like, assuming that NeuralHash is secure, it should either work well, that is produce an image similar to its input (because the hash is approximately unique) or not work at all (otherwise we would have found a collision), no middle-ground. Is there any thoughts/research on this?