Let's go Bananas with GenMedia — Guillaume Vernade, Google DeepMind
Summary
Guillaume Vernade, Developer Advocate at Google DeepMind, delivers a live workshop demonstrating a full generative media pipeline built on top of Google's suite of AI models. The session uses the public domain book 'The Wind in the Willows' (sourced from Project Gutenberg) as raw material to illustrate how Gemini, Imagen (Nano Banana 2), Veo, Lyria, and the Text-to-Speech (TTS) model can be chained together into a coherent multimodal content generation workflow.
Vernade begins by framing Google DeepMind's model strategy: the long-term vision is a single unified world model that ingests and outputs across all sensory modalities, but for release practicality the team ships specialized models (Imagen, Veo, Lyria, robotics, Gemma open models, AlphaGo-class research models). DeepMind ships on average every five days across all products, with the Gen Media stack receiving a new release at least monthly.
The core pipeline demonstrated: (1) The full book is uploaded to Gemini using the File Upload API, fed once into a chat session to exploit the large context window. Gemini acts as the prompt engineer for every downstream model. (2) Structured output (JSON schema with name + prompt fields) is requested from Gemini to produce character portrait prompts and per-chapter illustration prompts. (3) Nano Banana 2 (Gemini 2.1/2.5 Flash image preview model) generates portraits and chapter illustrations at resolutions from 520p up to 4K, with optional image grounding from web search. (4) Veo 3.1 animates chapter images into video clips by using the generated image as the first frame; a key improvement shown is asking Gemini to generate a video-specific prompt describing what happens in the seconds after the still image, rather than reusing the image prompt verbatim. Veo 3.1 Light costs $0.05/second ($0.40 for an 8-second clip), while the full Veo 3.1 model is more expensive. (5) Lyria (released two weeks prior to the talk) generates per-chapter music. The clip model produces 30-second tracks at $0.04 each; the full-song model produces up to 3-minute tracks at $0.08 each. All musical parameters — BPM, key, structure (intro/verse/chorus/outro), timing cues, and lyrics — are controlled entirely through the prompt, with no separate API parameters. Lyria is also shown generating songs from image inputs. (6) The TTS model (using voices Sulafat and Fenrir) reads character dialogue extracted from the book. A demonstrated trick achieves the illusion of four distinct character voices from two voice IDs by embedding per-character speaking-style instructions as parenthetical stage directions in the transcript, directing the model toward qualities like 'breathless and unbolstered' or 'long poetic pauses'.
Vernade introduces the Interactions API (in preview at time of talk, likely to become default at Google I/O) as a more efficient alternative to chat mode: it returns a stateful interactions ID that caches server-side context, eliminating the need to resend the full book on every turn, and supports forked conversation branches for parallel generation paths.
The Lyria Realtime model is highlighted as an underused gem: a predictive rather than diffusion-based model that streams music continuously and responds to new prompts mid-stream within approximately two seconds, functioning as a real-time DJ. A live demo of the 'Space DJ' AI Studio gallery app illustrates planetary prompt zones that blend musical styles as the cursor moves.
A notable architectural insight: all Gen Media models were trained on prompts written by Gemini, which is why using Gemini as a meta-prompt-engineer for Imagen, Veo, and Lyria yields unusually strong prompt adherence. The talk also discusses a three-tier service pricing structure: standard, Flex (50% discount with delayed processing), and Priority (2x price, fast-track processing). The total cost to run the full notebook is approximately $1, with video generation representing the bulk of that cost.
On European availability: Vernade acknowledges that all preview models are restricted to global endpoints (not EU data-residency endpoints), and that back-to-back model releases have repeatedly reset the GA counter. He identifies this as a personal P0 issue to resolve internally.
Key Takeaways
- •Use Gemini as the prompt engineer for all other Gen Media models (Imagen, Veo, Lyria) — the models were trained on Gemini-written prompts, so this yields significantly better adherence than human-authored prompts.
- •For video generation with Veo, ask Gemini to write a video-specific prompt describing what happens in the seconds after the first-frame image rather than reusing the image-generation prompt verbatim.
- •Lyria has no API parameters — all musical controls (BPM, key, song structure, lyrics with timestamps, language) are specified entirely in the prompt; this makes it unusually flexible but requires detailed prompts.
- •Use Lyria Realtime for dynamic applications such as game soundtracks that respond to player state, or interactive music experiences — it responds to new prompts within ~2 seconds mid-stream.
- •Achieve the illusion of multiple distinct TTS voice characters from just two voice IDs by embedding per-character speaking-style instructions as parenthetical stage directions in the transcript.
- •The Interactions API eliminates the need to resend large context (e.g., a full book) on every chat turn by caching it server-side and returning a reusable interactions ID — use it for multi-turn workflows to reduce latency and cost.
- •When generating per-chapter images with known character sets, pass only the reference images for characters actually appearing in that chapter rather than all character images, for better contextual consistency.
- •Veo 3.1 Light at $0.05/second (~$0.40 per 8-second clip) is suitable for fast prompt iteration; upscale to the full Veo 3.1 model only for final outputs.
- •Use Gemini's structured output (JSON schema) to get named, typed prompt objects from the model rather than free-form text, making downstream pipeline steps deterministic and easy to parse.
Transcript
[music] >> So, good morning everyone. Thank you for being here so early in the morning and to be able to be the few ones who pass security to to be actually here. Um I'm I'm Guillaume and I will talk tell you about uh Gemini in general. Um Like so this is this is my life as as Nano Banana sees it. Basically, I I joined Google 6 years ago. I used to be a video game producer before. I initially worked on Stadia, the streaming game companies that product [clears throat] that's uh that we killed as ...
Related Sessions
