← All speakers

Bio, Work & Ideas

Guillaume Vernade

Conference affiliation: Google DeepMind · 2026

Guillaume Vernade is a senior developer advocate at Google DeepMind specializing in Gemini’s image, video, speech, and music models. He helps developers build creative applications while pressing internal teams to simplify APIs, improve reliability, and address deployment constraints.

Vernade studied at the École Normale Supérieure and completed an MBA at the Collège des Ingénieurs. His early work included telecommunications, Google Glass and Android-watch prototypes, and roles as a product owner and Agile coach. At Ubisoft, he contributed to Assassin’s Creed Odyssey; he subsequently joined Google to work on Stadia, helping bring more than 30 games to the cloud-gaming platform, and later worked with Nest on the Matter connected-home standard.

After moving into developer advocacy at Google DeepMind, Vernade initially worked across the Gemini family before focusing on generative media. He contributes to the Gemini API Cookbook and maintains public software projects spanning home automation, presentation materials, and AI-assisted coding.

  • Multimodal creative workflows. Vernade designs pipelines that convert a public-domain novel into character portraits, scene illustrations, video, narration, and music. His book-to-media workshop shows Gemini producing structured prompts for specialized models, with retries for overloaded services and stateful interactions that reduce repeated context uploads.
  • Character consistency through selective references. Instead of accumulating every generated image in a sprawling conversation, Vernade identifies which characters appear in each scene and supplies only their relevant reference images. More elaborate productions can maintain front, side, and full-body views. His image-prompting guidance extends this approach to iterative editing, visual continuity, and layout control.
  • Cost-aware model selection. Vernade recommends generating inexpensive, low-resolution variations before upscaling strong candidates, matching model quality to task complexity, and using visual grounding when accuracy matters. His Nano Banana 2 guide also reflects his broader concern with enterprise permissions, regional data requirements, and the tradeoffs between the Gemini developer API and Vertex AI.
  • Inspectable AI-assisted coding. Vernade favors small, feature-specific files and explicit application logs, making unrelated agent edits easier to detect and failures easier to reconstruct. His Google AI Studio development guidance treats modularity and debugging as prerequisites for trustworthy generated software.

Read the topics behind these talks

2 conference talks

Key ideas

Scroll to read ↓

Follow video analysis, a bookshelf app, a book-to-media pipeline, and local Gemma agents through working results, failed assumptions, and live repairs.

  • From scientific Python to model infrastructure
    0:15 ↗
  • Which constraints will survive the next model?
    4:42 ↗
  • A toolbox of inputs, outputs, and application surfaces
    8:07 ↗
  • Turn a dinosaur video into a grounded table
    11:51 ↗
  • Bounding boxes versus a verification loop
    17:42 ↗
  • Specify an account-linked bookshelf catalog
    21:04 ↗
  • Change language while sharing a screen
    23:53 ↗
  • Explore a world generated as pixels
    28:56 ↗
  • Recognition succeeds before persistence does
    35:07 ↗
  • Choose the media model for the job
    43:11 ↗
  • Prepare a book-to-media pipeline
    51:27 ↗
  • Select the API surface and retain useful context
    54:20 ↗
  • Replace implicit image history with explicit references
    58:09 ↗
  • Animate the scene, then score the chapter
    1:03:01 ↗
  • Create distinct characters with two configured voices
    1:07:53 ↗
  • Convert the notebook without hiding the workflow
    1:13:40 ↗
  • Composition limits, service tiers, and capacity
    1:18:40 ↗
  • Fit the model to hardware you control
    1:24:26 ↗
  • Run skills and small coding loops on a phone
    1:29:42 ↗
  • Serve one local model to an agent workload
    1:32:47 ↗
  • Connect OpenCode, then test the generated game
    1:36:54 ↗
  • A game inside a game, and a screenshot into HTML
    1:44:43 ↗
  • Map language onto robot and development tools
    1:51:00 ↗
  • The final test is an interaction
    1:52:45 ↗

Key ideas

Scroll to read ↓

Guillaume Vernade builds a generative-media pipeline around The Wind in the Willows, showing how context, reference images and modality-specific prompts shape the results.

  • Why should changing an image model require changing the API?
    0:34 ↗
  • One multimodal ambition, separate releases
    2:44 ↗
  • Choose a model for each kind of output
    7:07 ↗
  • Turn a book into a bounded generation job
    9:43 ↗
  • Upload the book and establish a prompt contract
    15:33 ↗
  • Give the cast a consistent visual identity
    21:46 ↗
  • Move from conversation history to selected references
    26:10 ↗
  • A still-image prompt does not specify what happens next
    34:12 ↗
  • Compose related soundtracks without repeating the same song
    39:46 ↗
  • Make two configured voices sound like a larger cast
    47:53 ↗
  • Change the source, then change the interaction
    54:08 ↗
  • Availability and safety constrain the usable pipeline
    58:52 ↗
  • Use lyrics early and images as musical input
    1:03:05 ↗
  • Detailed instructions still need listening tests
    1:06:37 ↗
  • Navigate the music instead of submitting another song
    1:13:53 ↗

References