[ICCV2025] V2PE: Improving Multimodal Long-Context Capability of Vision-Language Models with Variable Visual Position Encoding - View it on GitHub