A new paper on arXiv presents CineRecon, a benchmark designed to evaluate how well frontier models can reconstruct reference videos as editable dynamic 3D scenes. The authors frame this as an important step toward controllable video generation for filmmaking, since such scenes act as visual proxies that can be modified and re-rendered.
The benchmark focuses on "agentic" reconstruction, meaning models are expected to perform the task in a goal-driven manner rather than through a fixed pipeline. By testing frontier models on this capability, CineRecon aims to measure progress in turning ordinary video footage into flexible 3D assets.
Because the abstract is brief, it does not include specific results or model rankings. The paper's contribution at this stage is the benchmark itself, along with the framing of editable 3D reconstruction as a key enabler for explicit creative control in video generation.