Adobe Firefly AI Model Development

Adobe Firefly prompt: A cinematic, photorealistic portrait of a confident, pensive Asian man in his 30s, standing in the rain against a neon-lit cityscape at night. He wears a futuristic black leather jacket slick with rain, droplets beading on his shoulders and damp hair. His face is turned slightly away from the camera, dramatically lit by a mix of soft neon lights and golden moonlight. The night sky glows navy and purple, with a small golden moon positioned just above and to the right of his head. Neon reflections shimmer on the wet pavement in streaks of pink, violet, and electric blue. The composition features cinematic lighting and film noir-inspired shadows, casting high-contrast highlights across his face and jacket. Captured with a shallow depth of field, the background blurs softly—rain-streaked signs glow abstractly, their symbols obscured by water and motion. The overall mood is atmospheric and futuristic, with a quiet intensity and melancholic tone.

To enhance the performance and creative range of Adobe Firefly's generative AI models across text-to-image and text-to-video capabilities, we needed to develop comprehensive datasets that moved beyond traditional commercial content limitations. The goal was to reflect diverse identities, authentic scenarios, cinematic aesthetics, realistic motion, and varied lighting conditions—all while ensuring commercial safety, legal compliance, and alignment with Adobe's responsible AI principles.

Collaborated closely with Adobe Firefly's computer scientists, researchers, strategic business development, content acquisition, and legal teams to understand existing datasets, performance goals, and evolving technical requirements. Led creative strategy across multiple AI initiatives, directing the Creative Development and Curation team to translate complex product needs into scalable creative frameworks and targeted mini briefs for crowdsourced contributors:

Text-to-Image Model Development: Developed creative briefs and visual quality frameworks that guided commissioned, acquired, and crowdsourced datasets. Expanded visual styles incorporating cinematic, editorial, documentary, and candid approaches while ensuring coverage of diverse lighting conditions, camera perspectives, and authentic representation to support comprehensive model learning.

Text-to-Video Model Development: Created content guidelines focused on cinematic aesthetics and realistic motion. Established frameworks for spatial depth and object-aware composition to improve model precision in localized generative editing workflows. This helped models better interpret human and object motion across varied perspectives to ensure high-fidelity environmental nuance.

Responsible AI Implementation: Developed content guidelines that ensured training datasets met Adobe's commercial safety and legal requirements across initiatives. Contributed to model evaluation and refinement, supporting the assessment of prompts and comparative evaluation of thousands of images to provide feedback to product teams for continuous refinement, reducing harm and bias in outputs.

The resulting datasets significantly enhanced Firefly's ability to generate content that was visually rich, technically sound, diverse, and aligned with user expectations. This foundational work:

Improved model capabilities to depict realistic motion, spatial depth, and environmental nuance across text-to-video generation

Set competitive differentiation by ensuring ethically sourced training data, distinguishing Firefly from competitors relying on scraped internet content

Enabled successful product launches, including the public beta release of Firefly's text-to-video tools in 2024, featuring generative b-roll and clip extension in Adobe Premiere Pro

Established scalable creative processes that extended across multiple content pipelines and AI model development initiatives