First-person rooftop parkour captured in Unreal Engine 5.8, made to train a whitebox → RGB video-to-video world model: the model takes a cheap, engine-free whitebox render of the scene as its control video and acts as the renderer. The whitebox is not rendered by Unreal. The courses are procedurally generated by a small Python package (pk, included under code/pk) that only uses boxes, cylinders and capsules. The same package renders the whitebox control frames with a lightweight OpenGL renderer, both for this dataset and at inference time. That lets you generate new worlds and control videos in milliseconds per frame without a game engine. Unreal is only used here to produce the paired RGB…