Baidu has been offering “AI” enhancement of its images for a while, and yesterday I bit the bullet, created an account, and trialled it.
It was a press image that I had in high-res, but Baidu’s image search, strangely, only had a low-res version of it. Could its “AI” turn it into a high-res image?
You can judge for yourself. It’s not bad but it makes the usual mistakes that “AI” does: it doesn’t know where there’s lettering, so it messes it up. The image, of the Roewe Clever microcar, looks reasonably good at first glance, till you see the front badge (it’s supposed to read Clever with the EV in the middle in caps and a larger point size) and the markings on the tyres.
I can see situations where this might come in handy, but for me it stresses the importance of sourcing an original image, even when it’s a rendering, rather than have software fill in the blanks for you. To me, the “AI” image looks incongruous. The use of the term hallucination isn’t that inappropriate when it comes to the “lies” large language models can concoct, and here the generated image looks like the result of a hallucination.
Judge for yourself: the top image is the original, the lower one is the “AI” enhancement. I’ve tried to crop them the same but they won’t be a perfect match. After all, I’m only human.
Here are some close-ups of the grille area, original on top, “AI” at the bottom. The bottom looks worse in a number of areas.
You may also like
