Video2LoRA: Parametric video internalization for VLMs

Arxiv CS.CV (computer vision)

Method reduces video token usage in vision-language models by internalizing video into LoRA parameters via a perceiver network. Achieves comparable performance to full-frame methods while using fewer tokens.

loravision-language

Arxiv CS.CV (computer vision)·Manan SuriSarvesh BaskarDinesh Manocha

7 days ago