Analyzing the Multilingual Performance of Text-to-Image Models

Explore the multilingual capabilities of text-to-image (T2I) models in this informative talk presented by Michael Saxon from UCSB at the USC Information Sciences Institute. Delve into the Conceptual Coverage Across Languages (CoCo-CroLa) benchmark, designed to measure the performance of T2I models across seven languages including English, Spanish, German, Chinese, Japanese, Hebrew, and Indonesian. Learn how this benchmark can be used to rank T2I models in terms of multilinguality and identify model-specific weaknesses, biases, and spurious correlations. Examine the qualitative analysis of success and failure cases for specific concepts, exploring how concepts are expressed differently across languages. Discuss the ethical implications of cross-lingual variations in model behaviors, ranging from culturally variable representations to potentially harmful biases. Consider the challenges of balancing bias removal with preserving cultural distinctiveness in T2I models. Gain insights into the future development of multilingual T2I systems and the need for targeted interventions to address language-specific performance disparities and demographic biases.

Analyzing the Multilingual Performance of Text-to-Image Models