Where Should Knowledge Enter? A Layered Framework for Knowledge Infusion in Multimodal Iterative Generative Model

The article presents a layered framework for knowledge infusion in multimodal iterative generative models, identifying four distinct intervention layers: surface, trajectory, latent, and parametric. It applies this framework to diffusion models, demonstrating that implementing multiple layers cumulatively can significantly reduce knowledge-violating outputs, achieving a 70.97% reduction compared to standard generation methods. This framework is crucial for practitioners as it provides a structured approach to enhance the reliability of generative models in safety-critical applications.