CINEMATIC
GENERATOR
THE_OBJECTIVE
Standard text-to-image models struggle to maintain consistent lighting, specific camera angles, and high-fidelity cinematic styling across multiple generations. The objective was to build a tailored pipeline that specifically generates professional-grade, cinematic portraiture.
CORE_SYSTEM_LOGIC
Implemented Low-Rank Adaptation (LoRA) to fine-tune a Stable Diffusion base model. I curated a specific dataset of cinematic stills, tagged them with specialized lighting and composition metadata, and trained the model to understand complex prompt weights.
UX_EXECUTION
Wrapped the underlying PyTorch inference engine in a clean Gradio web interface. This allows users to easily input prompts, adjust CFG scales, and tweak sampling steps without needing to interact directly with the command-line interface.