language:

• zh library_name: transformers license: apache-2.0 metrics:
• accuracy tags:
• multimodal pipeline_tag: video-text-to-text model-index:
• name: InternVideo2.5 results:
  • task: type: multimodal dataset: name: MLVU type: mlvu metrics:
    • type: accuracy value: 72.8 name: accuracy verified: true
  • task: type: multimodal dataset: name: MVBench type: mvbench metrics:
    • type: accuracy value: 75.7 name: accuracy verified: true
  • task: type: multimodal dataset: name: Perception Test type: percepTest metrics:
    • type: accuracy value: 74.9 name: accuracy verified: true
  • task: type: multimodal dataset: name: LongVideoBench type: longvideobench metrics:
    • type: accuracy value: 60.6 name: accuracy verified: true
  • task: type: multimodal dataset: name: VideoMME (w/o sub) type: videomme metrics:
    • type: accuracy value: 65.1 name: accuracy verified: true
  • task: type: multimodal dataset: name: LVBench type: lvbench metrics:
    • type: accuracy value: 46.4 name: accuracy verified: true

📕InternVideo2.5⚡

[📂 GitHub]
[📜 技术报告]

InternVideo2.5是一款基于**长上下文丰富建模（LRC）**增强的视频多模态大语言模型（MLLM，构建于InternVL2.5之上）。它通过密集视觉任务标注（采用直接偏好优化TPO）和自适应分层令牌压缩（HiCo）的紧凑时空表示，显著提升了现有MLLM在感知细粒度细节和捕捉长时态结构方面的能力。