I think this is not as complicated as it sounds; it's same as how NN classifiers aren't always built to do label to image, only image to label. Being able to do

I think this is not as complicated as it sounds; it's same as how NN classifiers aren't always built to do label to image, only image to label. Being able to do former through a parallel pipeline helps(that's like GAN), but that ability isn't required in training or in inference. Your OCR engine in the brain might generate "𰻞" for 'zh_hant_適' if reversed, doesn't mean it can't recognize the latter.