Video-Coder trains a coding agent to turn a text prompt into an executable Blender scene, then uses the rendered draft as structure-and-motion guidance for a separate video generator. With MiniMax-H3, it scores 0.67 on PhyGenBench versus 0.61 for MiniMax-H3 alone, and 78.19% on VBench-2.0’s physical dimensions—about 24 points higher than the base model.…
