VideoCoCo Uses Executable Blender Code as Chain-of-Thought, Lifting VBench-2.0 Plausibility From 52.18% to 77.88%

arXiv / HuggingFace Daily Papers·medium signalXBlueskyLinkedInCopy link

VideoCoCo (arXiv 2607.27380) splits physically-consistent video generation into two engines: a coding agent writes an executable Blender program specifying scene dynamics, then a video editor turns that deterministic simulation draft into photorealistic output. On VBench-2.0 it reaches 77.88% average plausibility versus the base generator's 52.18% — a 25.70-point jump — and on PhyGenBench it scores 0.558 average consistency against a 0.475 baseline, edging past Wan2.2-TI2V-5B's 0.544. The team also released VideoCoCo-3K, 3,000 agentically-curated draft/instruction/target triplets. The transferable idea is using a deterministic simulator as the reasoning substrate whenever a domain has hard physical constraints a diffusion model will otherwise cheerfully violate.