re: Help me understand AI

I would be extremely worried if my job was to create sketches of a suspect for police.

There’s not much intelligence left in the world so we have to invent it, sort of what silicon did for breasts.

"AI Generated Images" aren't AI, really....it's a sort of language model. On its own, it doesn't know what colors are, it doesn't know a square from a circle, it has no idea what "art" even is. But when you feed it vast amounts of data, it's able to match words with shapes/colors/etc. For example, say there are 7.5 billions instances on the internet of the word square used to describe an object...it determines that "square" must mean that particular object, which we humans know to be a square.

More broadly, the furthest we've reached with AI is large language models (LLM), such as ChatGPT or PaLM, etc. As others have mentioned, LLMs are only as reliable as the ingress data. For instance, you could ask an LLM what 10234552333542345x12341342334234 equals and it may not get it right. It's not an intelligent machine, it regurgitates what it's fed. If it was never fed an answer to 10234552333542345x12341342334234, it can't answer.

Artificial General Intelligence (AGI) is the big fricking deal where we get into serious philosophical and regulatory questions. These are the motherfrickers that can learn and function similarly to the human brain by applying logic, weighing options, etc. We've not reached AGI yet, but it's inevitable. Some scientists think it'll be a matter of years, while other experts believe we're decades away.

But what about a computer beating a chessmaster? Again, that's still not AGI. The computer doesn't know what "chess" is...it doesn't know a king from a knight from a pawn. Humans code the computer with board movements and define "winning". The computer is able to simulate hundreds of millions of different moves and different outcomes from each move in order to choose the most likely path to victory. It's not a good chess player, it's just able to process IMMENSE amounts of data very quickly.

quote:

---

1) Hands are complicated, lots of bones, joints, multi-axis movement.

2) In many pictures, the hands are not visible, or are partially hidden or intertwined with whatever they are holding. So the AI database of hands is much more limited than faces, which are prominent, and have people looking directly into the camera.

---

It’s largely a product of how diffusion models work. I posted a fairly lengthy explanation of how they’re trained on the first page so I won’t go through all of that again. But the main thing to remember with diffusion models is that they create images by removing noise in steps. Each step removes noise from the entire image. So it’s somewhat analogous to taking a blurry image and bringing it into focus. Everything comes into focus at once.

Contrast this with a human sketch artist who creates an image with lines - drawing the subject, then filling in the details. It’s an entirely different process.

The model doesn’t “know” anatomy. It just knows what final images should look like (oversimplified but again, I already described the process in more detail). There are computer programs that do know human anatomy on a 3D level and can recreate hands very easily. But that’s just not how diffusion models work.

As an aside, this is the same reason diffusion models struggle with text.

Another thing to consider is that hands in real photography are actually quite often pretty janky. As you said, they are complicated and as such create some weird perspective effects. If you’ve noticed fricked up hands in AI images to the point where you feel you can spot them, go start looking at some real photos on your phone or something. Seriously, you’d be amazed at how often hands look fricked up. So yeah that influences the model as well during training.

I suspect that at some point we will see different tools merged together so that you have a model that does understand anatomy to create the baseline for the picture, then a diffusion model to fill in the details. That actually already exists to an extent with control networks but they currently require manual input of poses, rather than being part of the generative process.