Veo 3: Visual Reasoning Through Chain-of-Frames. Large Vision Models. LLM, VLM. 3D U-Net backbone