google-research/kubric ? reverse-engineered prompt

Build me a tool that can generate synthetic videos of multiple objects moving around in a scene, with physics based interactions and realistic rendering. I want to be able to control how hard or simple each scene is, like changing the objects, textures, camera view, lighting, and number of things in the scene. It should also save useful labels for training machine learning models, like object masks, depth maps, optical flow, and other scene information.

Make it easy to run from a script or notebook, and include a simple example that creates a scene and outputs the generated data. If it helps, use Docker so it is easier to run consistently, and look up current docs online if you need to. I also want the project organized so it can support different kinds of dataset generation tasks later, not just one demo.