Anime characters are known for their big eyes, but this time it’s some tiny ones that are causing terror.

Japanese Twitter user Rampo Tamagawa (@tamagawa2525) likes video games, and especially old-school ones. He’s got a whole YouTube channel here where he uploads videos where he plays games like Sonic the Hedgehog, Street Fighter II, and landmark dating simulator franchise Tokimeki Memorial.

It’s an interesting channel if you enjoy taking another, closer look at games from yesteryear. However, Tamagawa ended up looking a little too close at Kaori Yae, one of his favorite Tokimeki Memorial girls, and made a terrifying discovery.

▼ Kaori Yae (non-terrifying version)

Now in fairness to Tokimeki Memorial developer Konami, its artists aren’t the ones responsible for the terror. Tamagawa was experimenting with an app called Remini, which uses AI to improve the quality of images. With the center line in the video below separating the “before” (left) and “after” (right), at first it looks like Remini did an admirable job getting rid of blurriness and adding definition, until Tamagawa zooms in close enough to reveal that the “improvements” made to Kaori include…

a face, with disturbingly realistic-looking eyes and lips, growing out of her chin!

That’s not the full freakishness extent, either. Remini also decided that Kaori’s nose would look better if instead of an air-passage orifice, it was yet another eyeball, giving her a total of five when you include the two in her chin.

Online reactions have included:

“Her true form has been discovered.”
“I’m a Kaori fan, so this is a huge shock for me (´;ω;`)”
“Yep, definitely gonna see that in my nightmares.”
“If I have a nightmare about this, I’m holding you responsible!”

But perhaps no comment has more succinctly captured the mood of Remini’s results as this:


So what happened here? Well, while programs like Remini are billed as being able to “improve” illustrations, they can only improve them in the sense that they guess what something is supposed to look like, and then alter the original to make it look closer to that supposed target. The potential problem is that an old, low-def game with anime-style illustrations, which are already inherently stylized, is a comparatively abstract form of visual expression. Since AI programs rely on what they’ve been able to observe and learn from other examples, it seems that Remini was more accustomed to human faces than anime girl drawings, and so its thought process wasn’t “Let’s make a slightly fuzzy anime chin a distinct anime chin,” but “Let’s make this really fuzzy face into a distinct human face.”

▼ Tamagawa spending some time with the original, non-monstrous Kaori

This could also explain why we’ve seen some pretty impressive results with AI programs being used to improve the smoothness of anime animation. In those cases, the program is being fed multiple starting images, including a beginning and end, to learn from, and is trying to create new images that are closer and closer to what it’s been given. That continually decreasing variance results in smaller increments of change, producing smother animation.

On the other hand, when it was given only a single still image, it appears that Remini didn’t have enough guidance as to which direction it was supposed to go in, and so it made a beeline for nightmare land.

Source: Twitter/@tamagawa2525 via Jin
● Want to hear about SoraNews24’s latest articles as soon as they’re published? Follow us on Facebook and Twitter!