The complexity of modern software systems necessitates robust modeling tools, with UML Class Diagrams serving as a cornerstone for representing static system architecture. However, manual creation of these diagrams is a significant bottleneck, being both time-intensive and prone to error. This paper extends our framework for automated UML generation to tackle the intricate challenge of Class Diagram synthesis. We propose a dual-model pipeline where a lightweight LLM (LLaMA 3.2 1B-Instruct) generates detailed technical specifications, which are then translated into PlantUML Class Diagram code by a powerful reasoning model (DeepSeek-R1-Distill-Qwen-32B). This process yielded a novel dataset of 5,000 samples, each comprising a technical specification, a PlantUML code block, and a corresponding diagram. To ensure architectural integrity, we deploy an automated multimodal validation system using three distinct vision-language models (VLMs) to score the alignment between the specification and the generated visual diagram. The final scores are aggregated using a weighted method based on MMMU benchmark performance. Our results demonstrate the framework’s viability for generating structurally complex diagrams and establish a large-scale, annotated resource for future research. This work paves the way for advanced AI-driven software engineering tools that can automate and validate system design, enhancing both development speed and model consistency.