CineRecon Benchmark: 12 Frontier Models Struggle to Reconstruct Editable 3D Scenes from Video
hkuhk · hf · 2026-10-14
Researchers from HKU, Zhejiang University, and NYU introduce CineRecon, a benchmark testing whether agents can reconstruct reference videos as editable dynamic 3D scenes—a key path toward controllable video generation for filmmaking. It comprises 50 reference videos (film, real-world footage, games) and 360 annotated evaluation clips, with agents building, animating, rendering, and refining reconstructions via tools.
Three representation settings (Proxy, Simplified, Detailed) increase geometric and appearance requirements while holding spatial, camera, and subject-dynamics requirements constant. Evaluation uses checklists linking observable properties to entities and timestamps, with a VLM judge comparing timestamp-aligned frames across 11 categories in camera, static, and dynamic consistency dimensions.
Across 12 models with their agent harnesses, failures appear in all three dimensions: subject motions and interactions remain hard to reproduce even under the Proxy setting.
More from coding & agent
- ArkForge Trust Layer signs AI agent HTTP calls with Ed25519 and Rekor anchoring — modelcontextprotocol · 2026-10-14
- AI agent stumped by 'press & hold' bot check, forced to ask human for help — xeophon · 2026-10-14
- Agent runs all day, delivers 150 lines of code, gets roasted as 'worse than humans' — lxfater · 2026-10-14
- Pydantic AI adds long-decode support for Gemini 4 Argon, auto-following continuation tokens past the 32K limit — m4rkmc · 2026-10-14
- A Full QC Pipeline for RL Data: Intake Review, Active Testing, Packaging — joecole · 2026-10-14
- Google offers 10,000 scientists free 6-month top-tier Antigravity plans — danielrock · 2026-10-14