MM1.5: Methods, Analysis & Insights from Multimodal LLM Fine-tuning

Haotian Zhang, Mingfei Gao, Zhe Gan, Philipp Dufter, Nina Wenzel, Forrest Huang, Dhruti Shah, Xianzhi Du, Bowen Zhang, Yanghao Li, Sam Dodge, Keen You, Zhen Yang, Aleksei Timofeev, Mingze Xu, Hong-You Chen, Jean-Philippe Fauconnier, Zhengfeng Lai, Haoxuan You, Zirui Wang, Afshin Dehghan, Peter Grasch, Yinfei Yang

🏛 Institutions: Apple
📅 Date: September 30, 2024
📑 Publisher: ICLR 2025 (Poster)
💻 Env
🔑 Keywords: model MM1.5 vision language model visual grounding reasoning data-centric analysis

TLDR

This paper introduces MM1.5, a family of multimodal large language models (MLLMs) ranging from 1B to 30B parameters, including dense and mixture-of-experts variants. MM1.5 enhances capabilities in text-rich image understanding, visual referring and grounding, and multi-image reasoning. The authors employ a data-centric training approach, utilizing high-quality OCR data and synthetic captions for continual pre-training, alongside an optimized visual instruction-tuning data mixture for supervised fine-tuning. Specialized variants, MM1.5-Video and MM1.5-UI, are designed for video understanding and mobile UI comprehension, respectively. Extensive empirical studies provide insights into the training processes, offering guidance for future MLLM development.

Open paper Edit on GitHub Report issue