The point is that it's not based on hallucination -- it's generated out of the authentic details provided from other images.
There's definitely a middle ground here that we perhaps don't have a good word for. E.g. what do we call a painting made by an artist who sat in front of the scene they depicted, vs. a painting made by an artist from their imagination? There's certainly some sense in which the first one was an "authentic" scene.
Yeah, except it's still absolutely vulnerable to hallucination. Look at the last set of images on "Limitations" page. The algorithm knows that there's a sign with text there, and it uses the original image to get the right letters in there, but it randomly reorders the letters rather than using the source image. "Real" and "authentic" is extremely misleading here.
That said, props to them for calling out the limitations so clearly. I really appreciate it when people are up front with the problems like that.
