Welcome to FLUX

How do we define realism, though? When I see an AI-generated image (or a 3d render), I can almost always tell it’s not real. Some are convincing at first glance, but with a closer look, there are usually details that give it away.

There’s some specific quality (idk what exactly) that makes it easy to distinguish AI or 3d renders from real photos. But even if you ignore that and look closely:

First image by Legolas: The right hand has six fingers, and the middle finger is smaller than the index finger. The moon also looks off—it seems unnaturally close, almost directly behind her, rather than in the distance.

Second image: The left hand has six fingers. There's a weird copyright text in the bottom left corner.*

Images by Frank:

The first image is quite good, but it features overly detailed collarbones (a common trait in AI images). The sweater has inconsistent textures on the top and bottom halves, and the ring looks misplaced.
The second image is better, but the skin looks overly smooth, which makes it appear artificial.

I'm sure there are other things too.

Imo, 3d is way better for achieving true realism because it has fewer limitations (the artist's skill being the biggest). Also, 3d allows you to get exactly what you want but AI is trial and error while also wasting exponentially more energy.

* There's also the entire debate about ethics, artwork being used to train models without the artist's permission (which is pretty much all models that are available) but since this is a pirate forum I won't go into that.

That being said, I really like the images though, they're great.

😀

I always found depth maps to be bette4 than openpose with flux

@frank22 The Henry Cavill Lora I'm using with Flux works quite well, and the prompt "Henry Cavill as a (insert fantasy... erm... profession here) with his shirt off" is producing reasonably consistent results, although I appreciate it may not be the Lora you're looking for.

I agree with you @frank22. AI can generate images way more realistic than anything Daz can render, with much less effort. But that's more of a DS limitation and not 3d in general. And DS is just a toy software.

What I wanted to say is that AI has inherent limitations (based on my understanding). What if you wanted to change the ring's design to a very specific shape. Surely, you can model it on your own but with AI it's just "prompt engineering" until it produces the desired output, which it might not be able to, at all.

Despite all the recent advancements, AI still can't make hands properly, simply because it doesn't "understand" anatomy, unlike a 3d artist; and hands are usually an insignificant portion of the images in the training data.

Maybe, there needs to be a model specifically trained on images of hands that can be used to fix those. Maybe something like this already exists, idk.

I'm sure you can't train AI models on AI generated stuff, it won't learn anything "new" from it, or even get worse. Also, more training data doesn't necessarily mean that AI will be able to produce better output, there are limitations to scaling.

Nothing is original, everything is derivative, execution is the only thing that matters.

Yes, but the thing is that you need human generated data to train AI and that's a limited resource. The web has AI art just about everywhere now. For that reason alone, the work of a person will always have a greater value for me than anything AI generated.

But after all, Daz and Flux are both just tools and you use whatever best fits your use-case.

(Apologies for hijacking the thread again, though.)

🙂