The GenMedia Model Lineup

▶ Watch (05:43)

Google DeepMind ships something from the GenMedia family more than once a month. Across all of DeepMind, releases average every five days. Vernade flagged 2025 as the year where agents stop being talked about and start being built:

“is going to be the the year of agents, the real one.” — Guillaume Vernade

The current lineup: Imagen (Nano Banana 2) for images at resolutions from 520 pixels to 4K, Veo 3.1 for video at $0.05 per second, and Lyria for music at $0.04 per 30-second clip. Nano Banana 2 adds image grounding, pulling live web images as references so the model knows what buildings and animals actually look like.

Choosing Between Gemini Developer API and Vertex AI

▶ Watch (16:21)

Three access paths exist for the Gemini models. The consumer app hides which model runs. Vertex AI exposes everything down to the data center but takes time to configure. The Gemini Developer API sits between: create a key in AI Studio, start calling models immediately, and swap to Vertex later without rewriting code because both use the same SDK. A new service tier, shipping that week, adds a Flex option at a 50% discount with possible delays and a Priority option at 2x cost for guaranteed throughput. Neither tier yet works with Veo.

Building the Book Illustration Pipeline

▶ Watch (18:55)

The demo notebook pulls “The Wind in the Willows” from Project Gutenberg, uploads it via the Gemini file upload API, then runs Gemini in chat mode to generate structured image prompts, one per character. Chat mode keeps every previously generated character image in context, so Imagen holds consistent fur colors and clothing across scenes. He walked through the book illustration Colab demo (18:55) live. For chapter illustrations, only the reference images for characters who actually appear in that chapter are passed as inputs, giving the model cleaner context.

Animating Scenes and Simulating Multiple Voices

▶ Watch (34:13)

Passing a chapter image to Veo sets it as the first frame. A first attempt reused the same prompt for both image and video, producing the wrong character speaking. The fix: ask Gemini to write a separate video-specific prompt describing the next few seconds of action from the starting image. See Veo first-frame video generation (34:50). For audio, the TTS multi-character voice demo (51:24) uses Sulafat for narration and Fenrier for all characters, steered into distinct speech patterns via style annotations embedded in the transcript text.

Lyria: Prompt-Driven Music Generation

▶ Watch (39:50)

Lyria has two batch models: a 30-second clip at $0.04 and a full-song model at $0.08, plus a real-time streaming model that responds to new genre prompts within two seconds. As Vernade put it:

“you everything is managed in the prompt.” — Guillaume Vernade

BPM, key, structure, lyrics, and language all go into the prompt text. For a structured song, write intro, verse, chorus, and outro blocks in plain text. The Lyria real-time DJ demo (55:54) shows live genre steering, shifting from post-punk to K-pop to chill mid-stream.

Q&A

New models are only available on global endpoints, not EU-only endpoints. Will that change? Vernade said the rule that every preview model lands on global endpoints first is unlikely to change, but Google is working to push models to general availability faster so EU-regional endpoints become available sooner. ▶ 58:54

Notable Quotes

So uh the short answer is no. Guillaume Vernade · ▶ 59:23

is going to be the the year of agents, the real one. Guillaume Vernade · ▶ 6:05

you everything is managed in the prompt. Guillaume Vernade · ▶ 43:39

Key Takeaways

  • Gemini writes prompts for Imagen and Lyria because all GenMedia models trained on Gemini-generated captions.
  • TTS style annotations in the transcript text simulate distinct character voices from a single voice model.
  • Lyria’s full-song model accepts structured section layouts, BPM, scale, and timestamped lyrics entirely in the prompt.