AI Video Production
Exploring generative AI video workflows from character development through final execution. Testing model capabilities, identifying limitations, and developing prompting strategies for commercial-quality outputs.
Character Creation
These characters were generated using Higgsfield's Nano Banana Pro model, prioritizing photorealism and consistency across iterations. The goal was creating portraits that felt authentic rather than obviously AI-generated. Both locked in within three iterations, establishing a visual foundation that could maintain consistency across different video models and prompts.
Model Testing & Iteration
Testing video generation across Kling, Veo 3, WAN, and Sora revealed consistent challenges: facial feature drift, anomalous background elements, furniture distortions, and the telltale "AI look" that breaks immersion. Each model required highly specific prompting covering foreground subject details, environmental context, lighting conditions, and background elements. Generic prompts produced unusable results. The successful outputs required granular direction at every layer of the frame.
Final Applications
Issey Miyake (Fashion): Testing fashion presentation against brutalist architecture. Evaluating the model's ability to handle abstract creative direction, maintain lighting consistency, render fabric accurately, and preserve character features across movement.
SJP Stash Perfume: Testing product integration and environmental realism. Focused on maintaining facial consistency, generating a believable loft apartment interior, and accurately rendering product packaging in-frame without distortion.