arxiv preprint - MG-LLaVA: Towards Multi-Granularity Visual Instruction Tuning | AI Breakdown | Podwise