Gen Media Is Ready—But Your Team Isn't Prepared for the Taxing Evaluation

Google DeepMind's panel on generative media centers on the launch of NanoBanana 2 Light and the Gemini Omni Flash API, which bring faster, cheaper image generation and video/editing capabilities to developers. The core claim is that video models are emerging as a foundational model for space-time understanding, not just a creative toy. The panel argues

that while language is a useful conditioning signal, it's inherently lossy for sensory details like audio, skin texture, or facial expressions. This frames the real challenge: the bottleneck is less about raw generation and more about evaluation, control, and bridging the gap between human taste and model defaults. The discussion reveals a key tension: models

that humans prefer in blind tests are often less realistic due to over-optimization for saturation and sharpness. This creates a governance and quality-control problem, as the default aesthetic of a model becomes a de facto choice made by engineers, not designers. The panel emphasizes that evaluation for generative video is underdeveloped—relying heavily on small human

panels and subjective side-by-side comparisons, unlike the mature evals for LLMs. This lack of reliable feedback loops is a major operational risk for teams deploying these models. For engineering leaders, the actionable signal is that the current generation of media models offers real workflow utility, particularly for ideation, iteration, and mid-fidelity content, but comes with