Retrieval-grounded LLM-based text-to-video generation and evaluation of synthetic construction accident cases

Construction safety training often relies on descriptive accident reports, which have limited ability to convey spatial configurations, causal sequences, and visual characteristics of hazardous situations. Real construction accident videos are also difficult to obtain because of privacy, ethical, and practical constraints. This study proposes a multimodal large language model (LLM) – based text-to-video approach for transforming narrative construction accident records into synthetic accident videos. The process integrates structured accident-scenario extraction, retrieval-augmented generation (RAG)-based prompt construction, reference image generation, text-to-video synthesis, and multi-level evaluation. Eight RAG-based videos and eight no-RAG baseline videos were generated and compared. Semantic grounding was assessed using embedding-based cosine similarity between Construction Safety Management Integrated Information (CSI) accident narratives and generated prompts. The RAG-based prompts achieved higher mean top1_cosine and topk_mean values than the no-RAG baseline prompts, with values of 0.640 versus 0.584 and 0.604 versus 0.548, respectively. Visual and temporal quality was evaluated using frame-based metrics, PSNR, SSIM, LPIPS, and temporal consistency. Object-level downstream utility was assessed using pretrained YOLO-based construction-hazard detection weights, achieving [email protected] of 0.850 and mAP50–95 of 0.684. The results demonstrate the feasibility of converting accident narratives into semantically grounded synthetic video resources for construction safety communication and downstream analytics.