No, this will most likely not happen in 2 years, because it's not a straightforward progression from the current state of the art.
Generating realistic video requires using 3D models, instead of the 2D models that Stable Diffusion and similar use now. To get beyond a crappy level, at least two problems have to be tackled:
1. More example data has to be available. There is an abundance of 2D video and still images, on which the current systems are trained. There is, however, not enough 3D footage.
2. The systems that generate 2D images now cannot easily scale up to 3D images. Point clouds would require ~1000 times the amount of data, and 3D meshes are so different from images, that current techniques won't work.
So, yes, it is possible that this will happen in two years time, but it seems highly unlikely.
What I deem possible in two years time is generating animations such as cartoons, in which a perfect understanding of a 3D world is not required. And this would still require a lot of human input to be anywhere near interesting.
Also, if you consult companies about this, and say that AGI is 4-5 years away, then I wonder if you are trolling, and I should possibly have refrained from writing this comment. If you are not trolling, please consult a book on AI, or any SF novel for that matter, from the 1960s.