A paper about synthetic datasets for vision-language models is accepted in CVPR 2026.