I wonder if this is doing basically the same thing as SORA/Kling, but with less compute/model size. It kind of reminds me of OpenAI's examples from their technical report: https://openai.com/index/video-generation-models-as-world-si...

Somewhere between "base compute" and "4x compute".

So maybe you "just" need to know how to create a certain type of diffusion transformer model and then train on a ton of videos, but with an adequate amount of compute. Which is probably a LOT for training and inference to get more realistic results.