Multimodal Generative AI: From Production Architectures to Multi-Aspect Evaluation Frameworks

Recent advances in multimodal generative artificial intelligence have transformed the synthesis and evaluation of multimedia content across text, image, video, audio, and ۳D modalities. This paper provides a concise, academically rigorous analysis of the core architectures powering this shift; including Generative Adversarial Networks (GANs), Diffusion Models, Autoregressive Transformers, and Multimodal Large Language Models (MLLMs); focusing on personalized education and digital health deployments. We analyze the limitations of standard statistical distance metrics (FID, FVD) and semantic alignment metrics (CLIPScore) in capturing localized visual anomalies, factual contradictions, and target domain utility. To bridge this gap, we propose a position-based selection framework that integrates multimodal judge models with qualitative error auditing, and introduce an operational governance matrix based on clinical risk and ethical alignment scores. Our findings stress that mitigating the proliferation of generative "slop" and "careless speech" requires a transition from purely quantitative metrics to multi-dimensional, human-in-the-loop evaluation frameworks that enforce transparency, safety, and regulatory compliance.

کلیدواژه ها:

Multimodal Generative AI ، Multimodal Large Language Models (MLLMs) ، Diffusion Transformers (DiT) ، Intelligent Agents ، Multimedia Learning ، Algorithmic Evaluation