Researchers have developed a new evaluation framework called Physics Question Scene Graph (PQSG) to assess the physical plausibility of videos generated by AI models. PQSG uses a hierarchical question-based approach, leveraging a vision-language model to identify violations of physical laws within generated content. The framework was validated using the FinePhyEval dataset, which includes human annotations, and demonstrated a higher correlation with human judgments than previous methods. The study also found that PQSG ranked closed-source models like Sora 2 and Veo 3 higher than Wan 2.1 in terms of physical realism. AI
IMPACT This framework could lead to more physically realistic AI-generated videos by providing better evaluation metrics.
RANK_REASON The cluster describes a new research paper introducing a novel evaluation framework for AI-generated videos.
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 3 sources. How we write summaries →
