MULTITEXTEDIT: Benchmarking Cross-Lingual Degradation in Text-in-Image Editing

In three linesMULTITEXTEDIT is a benchmark of 3,600 instances across 12 typologically diverse languages for evaluating text-in-image editing. Authors introduce a language fidelity (LSF) metric detecting script-level errors (missing diacritics, reversed RTL order). Evaluation of 12 systems reveals pronounced cross-lingual degradation, especially on Hebrew and Arabic.

Summary generated by Claude — human-verified