Upgrade your account to a Community Supporter account and remove most of the site ads.

Reply to thread

Message: <blockquote data-quote="EzekielRaiden" data-source="post: 9148828" data-attributes="member: 6790260">Missed this before, but I can give some answers here. I am not an expert at all, but I've been consuming Numberphile and Computerphile stuff that discusses this, as well as some things from other prominent math youtubers that go into the science of it (and actually have the accreditation to know what they're talking about.)More or less, pictures about humans usually involve very simple parts arranged in pretty simple ways. Faces are extremely important, featured in almost every image about people, and generally follow very simple patterns that are easy to recognize and reproduce. Likewise, limbs and torsos are quite simple to model. Clothing is a bit more complex, but it too tends to follow certain patterns just because...well, it's usually cloth draped over a body, so it has common physics. Armor, likewise, follows patterns. Both of them are more complex than faces or limbs, but not that much more.Hands, on the other (heh) hand? Good Lord. They're incredibly complex. They have to be at the end of arms, that's a blatant association. But fingers can take all sorts of very subtly different orientations and angles, and worse, they can easily blend together into a single blob or otherwise smudge out and lose detail. Hands are very rarely the central focus of an image, almost always at least somewhat peripheral, which means AI training algorithms will place less emphasis on those parts of the image. There are many, many, MANY photos of humans which don't even show hands at all, while there are comparatively almost none that show only hands and not faces.So...yeah. Hands have lots of fine detail that is easy to miss, they can take an enormous variety of positions which (to an AI) all look like completely different objects, and they're almost always at the periphery or at least away from the central focus of the image, which means the AI both collects less data and makes more approximations when asked to generate data.There's a reason we have sign language via hands, not via faces. Hands can carry enough linguistic information to communicate extremely subtle syntactic and semantic detail. Faces....not so much.</blockquote>

Verification