I do think AI sometimes has issues where it doesn't know certain rules that should apply to whatever scene you're doing. Humans have 2 arms, not 3. They have 10 fingers, not 14. Gravity is a thing. And so on. I assume that kind of thing will get better as the AI companies train their image generating models better so an individual user isn't going to have to try and figure out how to get the stupid AI they're using to stop giving humans 3 arms or dogs 2 heads or whatever.
Once we get 3D model generation that was trained Model Topology Data I think most of the issues will be resolved.
That Model Data is always accurate since most humanoids is based on a skeleton rig that is standardized and compatible with poses and animations.
3D Generation is also more about morphings things than generating new meshes from scratch, if it's trained on good topology it might actually do some miracles.
Something like Spore Creature Editor should also be possible.
Once we have the 3D models as well as a Simulation and Rendering Engine based on Procedural Materials that can also be generated, that can be fitted and together with a library worth of poses animations so we can fit it to a 3D scene based on that 2D image, that can then be feed back to that 2D AI generator to fix whatever problems they have.
So for people who cry about the AIs running out of Data to "Copy", once those 3D generators are made they can have an infinite amount of "Synthetic Data" to fix those issues specifically.
People forget that PBR stands for Physical Based Rendering, aka getting as close to simulating the real world, if we have that with Engines and Simulators then that is already effectively the same as the real world that can do that we can help the AIs figure things out.
Whatever you can do in Daz the AIs can do better.
Since there is an absolute need for that Synthetic Data and the Rendering and Simulation of that the big players are going to figure out 3D model and scene generation sooner or later.
Microsoft, Google and Unreal already have a head start in terms of training data for that.
Most artists nowadays break things down into pieces over multiple layers, they don't just draw the whole thing on a canvas.
The current AI image generators are the equivalent of drawing the whole thing on a canvas, of course they can get it wrong, humans can make mistakes and get it wrong while working without any reference images and directly with the paints without drawing things first.
Similarly AI generation will be broken down into multiple processes and pieces to fix this kind of issues.
People think AIs cannot have their "world view" in their minds, their Reference Model of how the World works, Bietch what do you think Daz and Unreal is? They can have as much Simulation as they want, what do you think real life Robots are trained on? Do you think they do trial and error in the real world?
