The increasing complexity of modern software systems underscores the need for robust, automated modeling tools. Among Unified Modeling Language (UML) artifacts, Sequence Diagrams play a crucial role in depicting dynamic interactions between system components. However, their manual creation remains time-consuming and error-prone. This paper extends our framework, focusing specifically on the synthesis of Sequence Diagrams. We introduce a two-stage pipeline that combines a lightweight language model (LLaMA 3.2 1B-Instruct) for generating detailed technical specifications, with a reasoning-enhanced large language model (DeepSeek-R1-Distill-Qwen-32B) to produce corresponding PlantUML code. This approach yields a novel dataset of 1,000 high-quality samples, each containing a technical description, PlantUML code, and the resulting diagram. To validate semantic and structural fidelity, we employ an automated multimodal evaluation system using three vision-language models (Qwen2.5-VL-3B, LLaMA3.2-VL-11B, Aya-Vision-8B). Each model independently scores the alignment between the textual specification and the generated visual representation. Final quality scores are aggregated using a weighted method informed by each model’s performance on the MMMU benchmark. Our experiments demonstrate that the proposed framework successfully generates architecturally consistent and semantically aligned Sequence Diagrams. Furthermore, our multimodal scoring system proves to be a reliable method for automated quality assurance, establishing a scalable benchmark for future research in AI-driven software engineering.