Compile the Event Logic Before Rendering

Abstract

Text-to-video models can produce individually plausible actions while violating the order, preconditions, or exclusions that make a procedure coherent. We propose an Agentic Vision architecture in which one multimodal controller reads instructions, writes an executable event plan, inspects Blender previews and generated video, and repairs localized violations across multiple rounds. We will test whether explicit event constraints survive neural rendering and improve all-constraint satisfaction at matched generation cost.

Keywords