VideoCoCo uses executable Blender code as chain-of-thought, lifting plausibility from 52.18% to 77.88%.

VideoCoCo splits physically-consistent video generation into a coding agent writing an executable Blender program specifying scene dynamics, then a video editor turning that deterministic simulation into photorealistic output. 25.70-point jump on VBench-2.0. The transferable idea: use a deterministic simulator as the reasoning substrate whenever your domain has hard constraints a diffusion model will cheerfully violate.