4 ms·

Good compression is quite complex and can go wrong in an unimaginable variety of ways. Remember when Xerox copiers would copy numbers incorrectly due to compre

• rasz 6y ago
  Xerox problem stemmed from simple replacement of "recognized" numbers with learned dictionary. Good implementation would use learned alphabet atlas as a supplement, encoding difference between guess and source image. That way even predicted 0 instead of 8 wouldnt be catastrophic with encoder filling in missing detail.