At this point the main thing I see missing to achieve more realism are better physics, global illumination and better facial animations. Models/textures are already there.
Now, it's animations that are too perfect, as you've pointed out. Even a perfectly trained athlete or ballet dancer still moves in slightly irregular and imperfect ways.
In this demo, most still frames could often easily be confused with a still frame from one of the films. Blown away, there. You're so right about the lighting, models and textures being virtually indistinguishable.
However - in motion, the effect is completely broken - only every few seconds - but enough to push it into uncanny valley territory.
I notice Nirobi (? sic, the Jada-Pinkett Smith's character) in the backseat seemed to be the most offensive in terms of unnatural animation.
I'm guessing - I'm hoping - ML models could be of tremendous use in terms of improving animation in video games? The problem is always that it is too unnaturally smooth. People just don't move like that.
Just as we got texture detail to the point of perfection of the unnatural - pores on the face, etc - we need to get animation to the perfection of the unnatural.
Even freaking same with the camera. After getting used to VR, the weirdly static, linear, unnatural camera movements in regular console games have aged like milk. Not sure why we are still treating cameras in 3D like it's 2001.