re: Here's What Popular Cartoon Characters Would Look Like IRL, According to AI

quote:

---

AI Olive Oyl >>>>>> Shelley Duvall

---

that's not much of a hurdle to clear though

Betty could get it.

Elmer Fudd looks a lot like Leslie Jordan

Need more pics of Velma

Damn Betty is a hottie.

Many of those are awesome almost scary how good.

Betty's definitely getting it

What’s up with all the freckles, AI?

Oh, and since when is Fred Flintstone the lead singer of RAMMSTEIN?

quote:

---

Can somebody who understands all this technology explain to me why AI can't do hands correctly?

---

It’s not that “AI” in general can’t do hands correctly.

It’s that diffusion models (the type of models behind basically all of the publicly available image generators) struggle with hands.

As for why diffusion models in particular tend to struggle with hands - it has to do with the way they work. Diffusion models start with an image made up of pure random noise and “denoise” them over multiple steps. It’s loosely analogous to looking for shapes in the clouds. So at step 0 you have pure noise, some number of steps later you have a really blurry image, and after the final step you have the final image.

The problem with this method is that the AI model isn’t trained to create a coherent “scene”; it’s trained to create a 2D image. It generally doesn’t have a good grasp of the entire 3D space. The images used to train the AI show hands in all sorts of shapes, sizes, perspectives, poses, etc… this confuses the model a bit because the model doesn’t really know what a “hand” is in the 3D, anatomical sense.

This is the same reason diffusion models struggle with text and with coherent backgrounds. For example, lines in the background (from bricks, doorways, ceilings, etc.) will often be offset on either side of the subject, because the model doesn’t really “think” about the room or structure in a 3D sense. Some diffusion models are better at various challenges than others, but they all struggle with these things to some extent.

This will eventually be solved, though, by using diffusion models in conjunction with other tools such as 3D wireframes, text generators, etc. (IMO). So in the future you might have a combination of methods where one model generates a scene using wireframes and chooses a camera perspective, and then another model adds textures to everything using diffusion. Just an example.