Text-to-Video Diffusion Model Resilience Study
The study reveals reliability risks in deployed T2V systems, particularly with bfloat16 format, affecting video generation quality and semantic accuracy.
The study reveals reliability risks in deployed T2V systems, particularly with bfloat16 format, affecting video generation quality and semantic accuracy.
这篇论文用Facet-Probe测试了18个主流多模态大模型,发现它们对输入顺序都很敏感,最好的模型也错13.4%,提醒我们模型可靠性还不是想象中那么好。