💡 This paper proposes MM-IFEngine, a multimodal instruction following data generation pipeline, accompanied by a 23k-scale training dataset and a dedicated benchmark, which improves MLLMs' instruction f