Researchers have introduced MPIE-Bench, a new benchmark designed to evaluate the ability of text-to-image and editing models to accurately depict multi-person interactions. The benchmark, comprising 2,500 video-mined editing triplets, focuses on anatomical plausibility and interaction accuracy, addressing common failures like fused limbs and interpenetrating bodies. MPIE-Eval, a new evaluation metric, scores contact-time geometry using mesh reconstruction, showing that current models struggle to excel in both anatomy and interaction simultaneously, often outperforming human judgment when assessed by vision-language models. AI
IMPACT This benchmark could drive improvements in AI's ability to generate realistic multi-person scenes, impacting creative tools and virtual environments.
RANK_REASON The cluster describes a new academic benchmark and evaluation metric for AI models.
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 3 sources. How we write summaries →
