Script-to-Scene: Cinematic Scene Generation via Fine-Tuning Stable Diffusion with DreamBooth and LoRA

The visualization of screenplays in the pre- production phase of filmmaking is a resource-intensive and subjective process. Traditional methods rely heavily on manual concept art, which is time-consuming and often lacks consistency. This paper proposes a ”Script-to-Scene” generation system that leverages Generative AI to automate the conversion of textual screenplay descriptions into high-fidelity, cinematic visualiza- tions. The proposed architecture utilizes Stable Diffusion (SDXL) as the core generative model, enhanced by a hybrid fine-tuning strategy. We employ DreamBooth to inject specific subject identi- ties (personalization) and Low-Rank Adaptation (LoRA) to learn and apply consistent cinematic styles without the computational cost of full model retraining. The system parses natural language scripts, extracts scene elements, and generates consistent visuals. Experimental results demonstrate the system’s ability to produce photorealistic, stylistically consistent scenes—such as specific character placements in atmospheric environments—validating its potential to streamline storyboarding and automated pre- visualization workflows.