Qwen-VL

• Paper: Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond
• GitHub Link
• Publisher: Arxiv
• Author Affiliation: Alibaba Group
• Functional Division
  • Understanding
  • Generation
• Design Division
  • Tool-using
  • End-to-end
• Input Modalities $\rightarrow$ Output Modalities
  (I: Image, V: Video, A: Audio, 3D: Point Cloud, T: Text, I_D: Document understanding, I_B: Output bounding box, I_M: Output segmentation mask, I_R: Output retrieved images)
  • I+T $\rightarrow$ T
• Model Architecture
  (Input $\rightarrow$ Modality Encoder $\rightarrow$ Input Projector $\rightarrow$ LLM Backbone $\rightarrow$ Output Projector $\rightarrow$ Modality Generator $\rightarrow$ Output)
  • Modality Encoder
    • I: ViT@448 initialized from OpenClip’s ViT-bigG
  • Input Projector
    • Cross-attention
  • LLM Backbone
    • Qwen-7B
  • Output Projector
    • None
  • Modality Generator
    • None
• Datasets Scale
  • Pre-training Stage
    • 1.4B
  • Instruction-tuning Stage
    • 50M