noise filtering.

[gmaxwell]

Try something like

                  xdelta = x - target_image
                  xdelta = xdelta - gaussian_filter(xdelta, sigma=1.5)
                  x = x - feedback*xdelta

before clipping. E.g. measure the error to the target image, high pass it, and feed that back in.

With feedback between 0.01 and 0.15 I often get pretty good results without breaking convergence.

One thing to do can be to converge with lower feedback then gradually increase it while the hash is still matching. It won't fix it if it settled into glitch art but the higher feedback can polish off the last of the noise.

Cheers.

[anishathalye]

Thanks for the suggestion! If you'd like to submit a PR, I'd be happy to merge it.

Otherwise, happy to merge a 3rd party PR as well, if someone else wants to implement this. Otherwise, I might get to it eventually.