Beyond Zero-Shot Business Process Diagram Extraction: Supervised Adaptation and Reliability Limits of Vision-Language Models

Business Process Model and Notation (BPMN) is widely used to represent complex workflows, yet many process diagrams survive only as images, limiting their use in computational analysis, model migration, and workflow tools. This study investigates direct image-to-JSON reconstruction with vision-language models (VLMs) through two complementary experimental stages. The first evaluates fourteen proprietary and open VLMs under zero-shot, optical-character-recognition-enriched, and prompt-ablation conditions using an aligned 202-diagram benchmark. The second adapts four open multimodal checkpoints using 3,478 English-translated BPMN image/XML pairs and low-rank adaptation. Canonical JSON targets are generated deterministically from the XML, and the adapted models are evaluated across three seeds for output validity, entity and relation extraction, prompt robustness, structural complexity, and external generalisation. To account for the loss of original scenario provenance, exact graph overlap is audited across the frozen dataset splits, and internal evaluation is distinguished from evaluation on a separate benchmark. Together, these stages provide a unified framework for studying zero-shot extraction and supervised adaptation while treating structural accuracy, output validity, dataset provenance, and diagram complexity as essential dimensions of BPMN reconstruction.