Training Video Models at Scale with Adobe Firefly

Explore the technical challenges and solutions behind scaling generative video models through this 22-minute conference talk from Fully Connected 2025 San Francisco featuring Adobe's Lior Shapira. Discover the key architectural decisions required to balance quality and efficiency in video generation models, including the implementation of diffusion transformers and overcoming VAE challenges. Learn about the complexities of curating diverse, temporally coherent datasets and developing responsible data strategies for training large-scale video models. Examine real-world infrastructure optimization techniques, mixed precision training approaches, and platform monitoring solutions used in production environments. Understand the unique debugging challenges in media models, from addressing confetti bugs to resolving blinking people artifacts. Gain insights into inference pipeline design, safety considerations, and deployment strategies for high-quality generative video models. Discover performance optimization techniques, load balancing strategies, and the integration of video generation capabilities into professional tools like Adobe Premiere Pro, drawing from practical experience in building and deploying Adobe Firefly Video GenAI at scale.