training-data multi-view consistent-character captioning ai-dataset