Unified Modeling Language (UML) diagrams are fundamental to software analysis and design, yet their manual creation consti-tutes a significant bottleneck, being labor-intensive and prone to incon-sistencies. While Large Language Models (LLMs) have shown promise in automating this process, a critical gap remains in the comparative analysis of generating different diagram types and in the scalable validation of their semantic correctness. This paper introduces a novel, endto-end pipeline that automates the generation and multimodal evaluation of UML Object and Sequence diagrams. Our workflow first employs a lightweight LLM to synthesize realistic software feature descriptions. Subsequently, a powerful reasoning-centric model, DeepSeek-R1-Distill-Qwen-32B, translates these descriptions into PlantUML code. The core contribution lies in our multimodal evaluation framework, which utilizes an ensemble of three distinct Vision-Language Models (VLMs) Qwen2. 5-VL, LLaMA-3.2-11B-Vision, and Aya-Vision to assess diagrammatic fidelity. Experimental results reveal distinct generation profiles: Object diagrams exhibit a high-competency, low-variability pattern, whereas Sequence diagrams present a polarized outcome, indicating greater modeling complexity. Crucially, our VLM based evaluation scores demonstrate a strong, statistically significant correlation with judgments from 155 human experts, validating VLMs as a reliable and scalable proxy for manual inspection. This work not only establishes the feasibility of a fully automated UML generation and validation workflow but also provides critical insights into the differing complexities of …