>>16626042Here's the way it works:
You basically collect a set of pictures, and you feed it to an algorithm that will identify certain patterns. It's not sentient at all, it just computes statistical patterns about shapes, colors, etc.
Then you can use it to spit back random images that use those patterns. So for example if you feed the algo with only pictures of trucks, it will only generate truck pictures.
Then, to generate pictures that look like the one you submit, what it does is generate tons of random pictures (using the patterns it has learned), and then it computes some score telling how close it matches the original (by seeing how many pixels are different in terms of color). It then iterates many times to improve the score, such that the random images look closer and closer to the one you give as input.
So back to your question, it simply does not know how to generate black characters, because they were not in the training sample. So it starts by generating white characters and various objects, and iterates until it matches the input picture.
