CineRecon: Benchmarking Frontier Models for Agentic Cinematic Reconstruction
CineRecon tests whether frontier agents can turn a reference video into an editable dynamic 3D scene.
The benchmark includes 50 reference videos and 360 annotated evaluation clips across film, real footage, and games. Agents are asked to construct, animate, render, and refine reconstructions under Proxy, Simplified, and Detailed settings. A vision-language judge checks timestamp-aligned frames across camera, static, and dynamic consistency categories. The authors report failures in all three areas, with subject motion and interactions still hard even in the Proxy setting. HF Daily Papers' note
The benchmark includes 50 reference videos and 360 annotated evaluation clips across film, real footage, and games. Agents are asked to construct, animate, render, and refine reconstructions under Proxy, Simplified, and Detailed settings. A vision-language judge checks timestamp-aligned frames across camera, static, and dynamic consistency categories. The authors report failures in all three areas, with subject motion and interactions still hard even in the Proxy setting. HF Daily Papers' note
score 5