Adobe Firefly | Text-to-Video
Creative Strategy for Generative AI | Visual Direction | Content Standards & Responsible AI
Challenge
Adobe Firefly’s text-to-video model needed high-quality, ethically sourced video datasets to support model training—ones that could teach the system to render realistic motion, cinematic aesthetics, and diverse lighting conditions, all while ensuring commercial safety and aligning with Adobe’s commitment to responsible AI. The challenge was to create content that elevated visual fidelity and met strict legal and ethical standards.
Approach
I led the creative strategy and guided my team in developing the creative briefs and content guidance that informed commissioned, crowdsourced, and acquired video assets. Working closely with Firefly product managers, engineers, legal, and business development teams, I translated technical model needs into clear, actionable direction focused on:
Cinematic style and visual tone
Lighting and atmospheric elements
Human and object motion across varied perspectives and camera angles
Content types and subject matter, including animation styles, b-roll, special effects, and atmospheric visual elements (e.g., smoke, fire, particles)
These categories were carefully selected to help the model better understand realistic motion, depth, and environmental nuance—enabling richer and more expressive video generation capabilities.
Impact
Improved the model’s ability to depict bodies and objects in motion across diverse camera angles and cinematic lighting scenarios
Helped ensure that all training content was commercially safe, legally cleared, and ethically sourced, setting Firefly apart from competitors relying on scraped internet data
Contributed to the successful public beta release of Firefly’s text-to-video tools in 2024, including features like generative b-roll and clip extension in Adobe Premiere Pro
“We’re transforming from a software company to an AI company… and doing it responsibly.”
— Alexandru Costin, Adobe VP of AI