Show the Futures the Evidence Supports

Abstract

A source clip can be compatible with multiple physical scene models that imply different outcomes after the same edit. We propose a multi-round Agentic Vision system that identifies these ambiguities, tests candidate scene programs against the source, and generates distinct counterfactual branches when their differences are consequential. We will evaluate whether branch-aware editing improves outcome coverage at fixed generation cost while preserving the source prefix and avoiding spurious branches on identifiable cases.

Keywords