(논문 요약) xGen-MM (BLIP-3): A Family of Open Large Multimodal Models (Paper)

핵심 내용

• Architecture: ViT (+perceiver resampler) and phi3-mini
• 데이터
  • DPP data: VLFeedback (data scored by GPT4-V)

실험 결과

• few-shot 성능은 Idefics2-8B 보다 떨어짐