Scholarlink

The exponential growth of large language models (LLMs)Large language models (LLMs) has opened up numerous possibilities for multi-modal AGIArtificial general intelligence (AGI) systems. However, progress in visionVision and vision-languageVision-language foundation models, critical for multi-modal AGI, lags behind LLMs. In this chapter, we discuss a large-scale vision-language foundation model (InternVL) InternVL, which scales up the vision foundation modelVision foundation model to 6 billion parameters and progressively aligns it with the LLM, using web-scale image-textImage–text data data from various sources. This model demonstrates state-of-the-artState-of-the-art performance across 32 generic visual-linguisticVisual-linguistic benchmarks, including visual perceptionVisual perception tasks such as image-levelImage-level recognition and pixel-level recognitionPixel-level recognition, vision-language tasks like zero-shot image/video classificationZero-shot classification and zero-shot image/video-text retrievalZero-shot retrieval, and integration with LLMs to develop powerful multi-modal dialog systemsMulti-modal dialog systems, which closes the performance gap from open-sourced MLLMsMultimodal Large language models (MLLMs) to commercial multimodal models like GPT-4VGPT-4V.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks

  • 章节

  • 发表日期：

    2026年

查看全文
  • Wenhai Wang,
  • Zhe Chen,
  • Yangzhou Liu,
  • Yue Cao,
  • Weiyun Wang,
  • Xizhou Zhu,
  • Lewei Lu,
  • Tong Lu,
  • Yu Qiao,
  • Jifeng Dai

摘要

The exponential growth of large language models (LLMs)Large language models (LLMs) has opened up numerous possibilities for multi-modal AGIArtificial general intelligence (AGI) systems. However, progress in visionVision and vision-languageVision-language foundation models, critical for multi-modal AGI, lags behind LLMs. In this chapter, we discuss a large-scale vision-language foundation model (InternVL) InternVL, which scales up the vision foundation modelVision foundation model to 6 billion parameters and progressively aligns it with the LLM, using web-scale image-textImage–text data data from various sources. This model demonstrates state-of-the-artState-of-the-art performance across 32 generic visual-linguisticVisual-linguistic benchmarks, including visual perceptionVisual perception tasks such as image-levelImage-level recognition and pixel-level recognitionPixel-level recognition, vision-language tasks like zero-shot image/video classificationZero-shot classification and zero-shot image/video-text retrievalZero-shot retrieval, and integration with LLMs to develop powerful multi-modal dialog systemsMulti-modal dialog systems, which closes the performance gap from open-sourced MLLMsMultimodal Large language models (MLLMs) to commercial multimodal models like GPT-4VGPT-4V.