Comprehensive Insights into Building Efficient and High-Performing Vision-Language Models

Rigorous experimental analysis of key design choices in vision-language models, leading to the development of Idefics2 - an open, state-of-the-art 8B parameter VLM that outperforms larger models on various benchmarks.