Nano Banana, Veo, and Lyria: Mastering the Google gen media stack

In this Google Cloud Next '26 session, Stephanie Wong interviews Khulan Davaajav, Product Marketing Manager for Gen Media, who delivers a live demonstration of Google's full generative media stack applied to a short cinematic film she created titled "Work From Home." The film tells a comedic story about snacking while working remotely, and serves as a vehicle to showcase how four distinct generative media models can be combined into a cohesive production pipeline.

Davaajav begins by clarifying the Gen Media umbrella, which encompasses Nano Banana (image generation and editing), Veo (video generation), Gemini Audio (transcription and text-to-speech), and Lyria (music generation). She emphasizes that model updates are released nearly every one to two weeks, signaling rapid iteration in the space.

For image generation, Davaajav used Nano Banana with highly specific artistic prompts referencing 33mm film, glossy highlights, halation, luminance, rounded 3D render geometry, and camera lens specifications. She explained her creative process of sourcing visual inspiration from platforms like Behance and Instagram, then using Gemini to help distill the artistic language into precise technical prompt vocabulary.

For video, she used Veo 3.1 Light, highlighted as Google's most cost-effective video model and capable of generating clips in under 60 seconds. She demonstrated two approaches: First-Frame/Last-Frame control, which animates between two defined keyframes with full control over start and end positions, and First-Frame-only mode, which allows the model to infer the remainder of the scene (e.g., panning around a character). She advised developers building video applications to encode common creative prompts (such as "dolly zoom" or "360-degree camera movement") as application-level features rather than requiring users to type cinematography terms.

For music, Davaajav demonstrated Lyria 3 Pro's timestamp-based prompt control, directing musical tone changes at precise moments — a chill 9AM vibe, a silence during a dance sequence, and a lullaby during a sugar crash scene triggered at 0.25 seconds. She noted that all diegetic sound effects in the video (yawning, turning a radio dial) were generated directly by Veo 3.1 Light rather than sourced separately. Gemini's multimodal understanding was used to automatically write the Lyria prompt by analyzing the video frame by frame.

For voiceover, she showcased Gemini 3.1 Flash text-to-speech, which offers over 200 expressiveness tags (e.g., [positive], [panicked], [laugh]) to control emotional delivery. Voice style parameters including accent (British casual), tone (comedic), and narrative style can be layered with the expressiveness tags for highly nuanced AI narration.

The session concluded with a live preview of the Gemini 3.1 Flash Live Avatar feature, which demonstrated real-time audio-to-audio interaction with accurate lip-syncing and live Google Search integration — the avatar correctly reported live Las Vegas weather during the demo. This was distinguished from static AI avatar products as a fully interactive, real-time system.

Looking ahead, Davaajav identified two developments she is most excited about: World Models (specifically Genie 3), which shift the creator from asset generation to operating as a first-person camera operator inside a fully simulated environment, and reducing generation latency to keep creators in a flow state during iterative work.