Google's FVScore: neuro-symbolic video evaluation that outperforms holistic MLLM judges
google · hf · 2026-11-19
Google Cloud AI Research and NUS introduce FVScore, a specification-driven neuro-symbolic framework for evaluating video generation.
- Existing evaluators rely on holistic MLLM judgments or vision-language metrics, which align poorly with humans and can't trace specific temporal violations
- FVScore compiles each video into explicit propositions (prompt requirements) and implicit propositions (unstated physical, temporal, cross-modal constraints), verifies them with specialized perceptual tools, and evaluates confidence traces with soft temporal operators inspired by Linear Temporal Logic
- A Perception-Aware Calibration step keeps only perceptually significant violations, producing timestamp- and evidence-grounded reports
- Across public and complex multi-event benchmarks, FVScore beats holistic and agent-based evaluators in human correlation while giving more accurate, interpretable feedback
More from Multimodal
- Kezek splits mono two-person conversations into per-speaker tracks for full-duplex dialogue training data — manifestro · 2026-11-19
- LTX-2.3 long-form lessons: CFG=1 kills negative prompts, brightness drifts per clip — Chance-Thanks-1278 · 2026-11-19
- 6-minute Hubble documentary made with Claude and Jimeng CLI workflows — xiaohu · 2026-11-19
- Seedance 2.5 4K hands-on: AI video finally loses its plastic look — xiaohu · 2026-11-19
- Eerie AI-generated portraits: neon grimace faces and glowing-eyed shadows — azed_ai · 2026-11-19
- Vaporwave Urban Mirage: a reusable AI image prompt template with sample outputs — LudovicCreator · 2026-11-19