CULTURESCORE: Evaluating Cultural Faithfulness in Video Generation Models
Researchers have developed CultureScore, a new framework to evaluate the cultural faithfulness of video generation models. This framework assesses identity representation, contextual accuracy, and behavioral norms across 10 countries. The study found that current state-of-the-art models fail to generate culturally accurate videos, with the best performer achieving only 56.8% on the CultureScore. Human evaluators ranked models based on cultural faithfulness, which often contrasted with purely visual quality metrics. AI
IMPACT Highlights a critical gap in video generation models, emphasizing the need for culturally aware AI development and evaluation.