MLLMs trained primarily on real-world data struggle to interpret stylized animated videos and often fail to recognize unseen characters. During video reframing, key characters may also be cropped out of the frame. To address these issues, we propose a training-free short-form transformation pipeline that uses character images as visual queries. The pipeline combines MLLM-based scene extraction, visual-query-based object detection, and Adaptive Zooming to reduce object loss and cropping errors.

Kanghee Lee received a B.S. degree in Business Analytics from Vrije Universiteit Amsterdam in 2022 and an M.S. degree in AI Imaging from Chung-Ang University. He is currently pursuing a Ph.D. degree at Chung-Ang University, with research interests in domain generalization, vision-language models, and video large language models.