VideoCoCo: Code-as-CoT for Physically-Consistent Video Generation via an Agentic Dual-Engine System
VideoCoCo turns Blender code into the intermediate reasoning step for video generation.
A coding agent writes an executable Blender program from the prompt, using it to define both the scene and its motion over time. The system renders that deterministic draft, then a generative video engine edits it into a photorealistic result. The authors also introduce VideoCoCo-3K to train the editor on simulated draft, instruction, and target triplets. They report gains over OmniWeaving on PhyGenBench and VBench-2.0, with the best average score on both benchmarks. HF Daily Papers' note
A coding agent writes an executable Blender program from the prompt, using it to define both the scene and its motion over time. The system renders that deterministic draft, then a generative video engine edits it into a photorealistic result. The authors also introduce VideoCoCo-3K to train the editor on simulated draft, instruction, and target triplets. They report gains over OmniWeaving on PhyGenBench and VBench-2.0, with the best average score on both benchmarks. HF Daily Papers' note
score 5