AI Engineer Europe 2026

Let's go Bananas with GenMedia

Loading slides…

Read the talk

Let's go Bananas with GenMedia

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.

From a talk by Guillaume Vernade

Before you start: Basic Python and familiarity with model API requests will help you follow the notebook; running the media-generation steps also requires appropriate API access and billing.

Why should changing an image model require changing the API?

An image-generation application ought to be able to change models by changing a model name. Yet Imagen and Nano Banana exposed different APIs, making that simple substitution more work than it needed to be. This is the practical problem behind Guillaume Vernade’s introduction to generative media: powerful models still need interfaces developers can use coherently.

His career began in video-game production, continued through Stadia and brought him to DeepMind. Vernade approaches that problem as a developer advocate: providing documentation, code samples, demonstrations, skills and prompt guides, while carrying developers’ problems back to the teams building the models. The illustrated career slide makes that background visible; the API complaint makes its relevance concrete. He jokes that Imagen’s apparent disappearance as a brand might have won the argument for him by default.

Speaker portrait and name beside an illustrated landscape connecting academia, video games, and Google and DeepMind.Open full source frame
Guillaume Vernade’s illustrated career journey.
0:340:47
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:34 · section reference included

One multimodal ambition, separate releases

Images, sound and video are ordinary parts of the world an application must understand. Vernade describes a world model as accepting as many of these inputs as possible—including sensor data—and responding through multiple modalities. The ambition is an encompassing model; shipping separate image and video models creates smaller release boundaries, reducing the chance that improving one capability breaks another.

His account of early Gemini releases illustrates a different boundary: a model’s underlying capability and the behavior exposed to users can diverge. He recalls image understanding being suppressed in an early release, then residual refusal behavior appearing in later versions until the move to Gemini 2.0. That recollection should not be read as a complete public release history: Google’s December 2023 Pro Vision announcement already offered image input before Gemini 1.5. The proposed connection between earlier suppression and later refusals remains Vernade’s explanation.

The portfolio extends beyond media to vision-dependent robotics, agents, open models and specialist research systems. Vernade corrects the slide’s Gemma 3 label to Gemma 4, which he describes as newly released, and names AlphaEvolve, AlphaGenome and WeatherNext. He expects the year’s agent work to shift from discussion toward building. He reports GenMedia releases more often than monthly and a DeepMind-wide average of one release every five days, with smaller features arriving more frequently. That pace explains why SDKs, examples and developer guidance are continually moving targets.

Slide titled “Beyond Gemini models” with illustrated panels for image, video, music, robotics, agents, Gemma 3, and additional models.Open full source frame
Beyond Gemini: images, video, music, robotics, agents, and open models.
2:443:24
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

2:44 · section reference included

Choose a model for each kind of output

Nano Banana 2 adds image sizing and aspect-ratio choices, search grounding and web-image grounding. The last capability supplies visual references for subjects such as architecture and animals, where knowing what something looks like matters as much as knowing its name. Vernade also describes restrictions on using images of newer buildings, so grounding is not available uniformly for every requested subject.

Veo 3.1 Lite offers a cheaper way to iterate on video prompts before moving to a more expensive result or upscaling. Vernade estimates $0.05 per generated second and about $0.40 for a video. The pricing table checked in August 2026 supports $0.05 per second specifically for 720p video with audio; $0.40 therefore corresponds to eight seconds under those conditions.

Lyria produces completed musical outputs: clips or full songs. Lyria RealTime instead keeps generating music while a session runs and accepts new prompts that steer the ongoing performance. Vernade describes it as predictive rather than the prompt-to-completed-output diffusion process he contrasts it with. He describes roughly two-second responsiveness when steering Lyria RealTime, rather than presenting a measured latency guarantee. The interaction resembles a DJ changing the mix while the music continues.

7:077:27
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

7:07 · section reference included

Turn a book into a bounded generation job

The workshop now moves from slides to a book-illustration notebook that participants can run alongside the demonstration. Vernade estimates the notebook run at about $1, with video generation accounting for most of the expense. Skipping video makes participation cheaper without removing the image workflow.

The pipeline starts with a book from Project Gutenberg. Gemini reads the book and writes prompts; the media models turn those prompts into character portraits, chapter illustrations, videos, music and narration. The Gemini Cookbook example combines capabilities that the repository also teaches separately through quickstarts and feature examples.

Colab notebook introduction titled “Illustrating a book using Nano Banana and genMedia models,” with an Open in Colab button and bullets for Veo, Lyria, and TTS.Open full source frame
The book-illustration notebook introduces images, video, music, and narration.

Setup requires an SDK recent enough to include the demonstrated music support and an API key from AI Studio. The main workflow uses paid access. For image-only participation, Vernade suggests the original Nano Banana model’s then-available free tier. His SDK guidance changes as he speaks, so the useful requirement is support for the selected models rather than an inferred minimum version.

The client loads the API key and enables automatic retries to handle overload, particularly when US traffic rises. Vernade describes five retries with a two-second delay. He selects Gemini 3.1 Flash Image Preview for Nano Banana 2, switches the text model from Gemini 2.5 Flash to Gemini 3 Flash, and chooses Lyria’s clip model and a Pro text-to-speech model.

Bound the workload before generating assets. The notebook has a payment-confirmation checkbox, although Vernade notices that he accidentally left it enabled by default. It also limits the number of characters and chapters. Those limits control both spending and waiting time: a whole book can expand into many independent generation requests.

9:439:52
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:43 · section reference included

Upload the book and establish a prompt contract

The source is Kenneth Grahame’s The Wind in the Willows. After downloading it from Project Gutenberg, the notebook uploads it with client.files.upload. That call leads to a useful distinction between Google’s access surfaces.

SurfacePrimary tradeoff
Consumer Gemini appsEasy to use; limited model and parameter control
Vertex AIInfrastructure, location and access control; more setup
Gemini Developer APIAPI-key onboarding; responsibility for protecting the key
AI StudioInteractive model experimentation

Vertex AI and the Developer API share an SDK, but their surrounding storage workflows differ. The Developer API’s file upload hides bucket creation and access-control configuration, making the uploaded file available to the model. A leaked API key remains a direct misuse risk.

The notebook creates a text chat so later requests retain the book and prior instructions. It also asks for structured output: each generated prompt has a name identifying its character or chapter and a prompt string. A Python schema makes that contract explicit:

python

from pydantic import BaseModel

class NamedPrompt(BaseModel):
    name: str
    prompt: str

The chat configuration requests JSON matching the schema, so downstream generation can consume fields instead of extracting instructions from free-form prose.

For the live run, Vernade adds Priority service, then tells participants to remove it if they want to save money. His recording-time comparison is:

TierPrice described in the workshopScheduling behavior
StandardStandard priceShared queue
Flex50% discountRequests may wait minutes
PriorityTwice the standard pricePreferential scheduling

These are his workshop descriptions, not a universal pricing or latency contract across models. He is unsure about Veo support and says Lyria does not yet support the option. The initial chat message then supplies the whole book while telling the model to wait for further instructions.

15:3315:48
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

15:33 · section reference included

Give the cast a consistent visual identity

Leaving style unspecified repeatedly produces similar defaults, so Vernade supplies a colorful building-block style. Colab’s form annotations make that value easy to edit. System instructions then constrain the composition: no book-cover titles and no multi-panel images. Those rules address a specific failure mode—when the model knows it is illustrating a book, it may produce a cover instead of a portrait or scene.

Colab shows the style “a colorful building block style,” instructions excluding text, borders and panels, and the heading for generating character portraits.Open full source frame
Style and image constraints precede character generation.

Gemini first describes the main characters. An adult-only restriction remains in the prompt from an earlier version of the example; Vernade says it can now be removed. He distinguishes generating new images containing children from editing existing images containing children, which he describes as restricted in Europe at the time. That policy history explains the leftover instruction without making it part of the book pipeline’s essential design.

A separate image chat receives the character prompts. It sets the image response modality, aspect ratio, system instructions, style and the demo’s Priority option. Keeping it separate avoids mixing the text prompt-writing history with image outputs, while retaining earlier images for visual continuity. Mole and Water Rat appear first, followed by Toad—conspicuously larger than his car—and then Badger and Otter.

Sequential generation gives each turn the images produced before it. Independent asynchronous calls could reduce waiting, but would not preserve that same ordered dependency. The notebook also appends the generated character images to an array for reuse. Vernade calls this demonstration-level storage; the important asset is the saved character reference, not the array itself.

21:4622:03
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

21:46 · section reference included

Move from conversation history to selected references

With Otter generated, the next request asks for one illustration prompt per chapter. It explicitly excludes multi-panel compositions and repeats character descriptions even though image references are available. The first implementation relies on chat history to carry appearance forward. That history has a cost: the SDK sends previous messages back on every call, including the book context, which can make repeated prompt generation slow.

The Interactions API offers a different context mechanism. A response supplies an interaction ID that a later request can reference, letting the server recover prior context. The same earlier interaction can also support branches: write lyrics, then branch into cover-image generation and song generation. Vernade’s tentative two-day retention answer is not a reliable storage contract; documentation checked in August 2026 specifies default retention of 55 days for paid interactions and one day for free interactions, with shorter paid retention configurable. His suggestion that Interactions might become the default at I/O is a forecast, not part of the demonstrated interface.

Vernade also points to automatic caching for reused context, while noting that ordinary API calls can benefit from caching too. The resulting illustrations show the riverbank, the road with Toad and the snowy forest. Their character consistency initially depends on the image chat retaining the earlier portraits.

A more explicit approach expands the structured output to include the characters appearing in each chapter. The application looks up their saved images and sends only those references with a unary generate_content request. The river scene uses Mole and Water Rat; the road scene’s requested cast includes Toad, Mole, Water Rat and a gray horse. A dictionary makes the lookup precise:

python

from pydantic import BaseModel

class ChapterPrompt(BaseModel):
    name: str
    prompt: str
    characters: list[str]

def chapter_contents(chapter: ChapterPrompt, images_by_character: dict):
    references = [
        images_by_character[name]
        for name in chapter.characters
    ]
    return [chapter.prompt, *references]

Here, the dictionary holds previously generated image objects. A missing character raises a lookup error instead of silently substituting an unrelated image. The request’s context is now an explicit selection of assets, rather than the accumulated conversation.

For a larger production, Vernade would generate multiple views of each character—portrait, full body, side and rear—and select the view that matches the planned composition. That is a proposed improvement, not what this run establishes. The explicit-reference results look broadly similar to the chat-based ones. One snowy scene seems to show Water Rat attacking Mole; inspecting the prompt reveals that the intended action is a rescue. Preserving appearance does not guarantee that the depicted action reads correctly.

26:1026:25
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

26:10 · section reference included

A still-image prompt does not specify what happens next

The next stage passes the final chapter illustration to Veo 3.1. In this image-to-video path, the supplied image becomes the first frame; the first attempt reuses the illustration prompt. Vernade emphasizes how much that starting frame determines the video’s subsequent possibilities. Answering a model-selection question, he identifies plain Veo 3.1 as the quality-oriented option and describes Fast and Lite as smaller, faster variants doing less generation work. He chooses the larger model for the demonstration while leaving cheaper options available to participants.

The generated rescue clip includes sound and dialogue, but the wrong character speaks. The still-image prompt described a scene without adequately directing subsequent action or assigning speech. The image supplied a starting composition; it did not supply a complete temporal script.

Vernade adds another prompt-writing step:

  1. Give Gemini the chapter image.
  2. Ask for a Veo prompt describing the next few seconds after that image.
  3. Generate video from the same first frame using the new continuation.

The resulting prompt has Water Rat lower his silver pistols and reassure Mole. Mole exhales white plastic vapor in relief, then the pair begin walking away. The language retains the book’s old-fashioned tone while carrying forward the building-block material and clothing details.

Vernade judges the second clip better, although it contains no conversation. This is a more directed continuation, not a demonstration that every desired audiovisual detail is now controlled. He also warns against regenerating video endlessly: iteration costs become significant when a vague prompt invites repeated trial and error.

34:1234:27
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

34:12 · section reference included

Compose related soundtracks without repeating the same song

For music, the workshop uses Lyria’s faster clip model. Vernade estimates $0.04 for a 30-second clip and $0.08 for a full song of up to three minutes. Gemini again supplies the prompts, this time asking for instrumental music that shares an identity across chapters while emphasizing what makes each chapter different.

Vernade attributes Gemini’s usefulness as a media prompt writer partly to training: he says Gemini-written prompts contribute substantially to GenMedia training data. He also describes internal prompt rewriting and recommends giving the model detailed direction instead of a one-line request. The playback provides a concrete comparison: pastoral music for spring and flowing water, a more adventurous treatment for the open road, and tension for the dark forest.

Lyria’s musical controls live largely in the prompt:

  • Instrumentation and feel: Specify instruments, how they play, scale and BPM.
  • Duration and development: Describe when the arrangement should change.
  • Song structure: Identify an intro, verse, chorus or outro; the chorus indicates material intended to recur.
  • Lyrics: Supply the words or ask the model to invent them.

Vernade changes the chapter request from instrumental tracks to songs whose lyrics explain the chapter events.

He reports that the 30-second music model generates in a few seconds in this workflow, while resending the book and writing prompts takes longer. The resulting songs describe Mole abandoning his cleaning and Toad dreaming of the road in a yellow-and-red caravan. Their themes remain similar to the earlier music: chat history helps preserve continuity, but also encourages the prompt writer to reuse familiar descriptions and musical ideas.

Lyrics are part of the generation prompt, and individual passages can be associated with moments in the song. The returned lyrics also carry timing information, making synchronized text or a karaoke interface possible. The final chapter song narrates Mole’s fear in the snowy wood and Ratty’s arrival to rescue him. The same chapter has now become an image, a moving scene and a musical retelling.

39:4639:58
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

39:46 · section reference included

Make two configured voices sound like a larger cast

NotebookLM’s conversational audio offers a familiar example of two configured voices. The workshop explores how delivery instructions can make those voice slots support more perceived characters. Gemini extracts a passage from the book and rewrites it as a play, with narrator and character lines and a consistent speaking style for each character. The passage begins with the description of Water Rat’s small ears and silky hair and ends at the specified phrase about ears in the air.

The narrator gets one voice; the other characters share another. Parenthetical directions distinguish them through long poetic pauses, breathlessness, stuttering, whispers or emotional intensity. Repeating a character’s delivery across lines gives the listener a stable cue even when the underlying voice is shared.

Colab code cell requests dialogue formatted as a play, with narrator and character lines and consistent character speaking styles.Open full source frame
The dialogue prompt specifies a distinct speaking style for each character.

The TTS request needs an explicit reading instruction. Vernade had lost time by supplying text without first asking the model to read it. He maps the narrator to Sulafat and the shared character slot to Fenrir. The corresponding prompt assembly can remain simple:

python

def reading_prompt(play_text: str) -> str:
    return (
        "Read this text. Follow the parenthetical delivery directions "
        "and keep each character's speaking style consistent.\n\n"
        + play_text
    )

voice_assignments = {
    "narrator": "Sulafat",
    "character": "Fenrir",
}

The voice assignments belong in the speech configuration; the play text carries the local performance directions.

Playback follows Mole and Rat meeting and boarding the boat. The characters sound distinguishable, although Vernade notices that both are speaking slowly; he also describes TTS generation itself as slow. Delivery steering creates character differences here without reliably controlling the pacing.

For a larger application, Vernade would keep actual character names in the transcript and maintain a separate baseline prompt for each character. A line can then add local excitement, fear or urgency without replacing that baseline identity. He also notes that the requested accent did not come through. Irish, Singlish and German accents are possible additional directions to explore, rather than successful results demonstrated here.

47:5348:05
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

47:53 · section reference included

Change the source, then change the interaction

The book’s text is only one possible input. A large-context multimodal model can also work from an audiobook or video. Vernade starts another illustration run using The Adventures of Chatterer the Red Squirrel, this time requesting a futuristic science-fiction utopia with saturated neon lights. The narrative source and visual treatment are independently adjustable.

While that runs, he opens a Lyria RealTime example in AI Studio’s GenMedia gallery. The music begins with post-punk and neo-soul, then changes as he asks for more K-pop, more drums, less post-punk and finally a calmer mood. These are updates to an ongoing performance rather than requests for separate finished songs.

His game-production background suggests a natural application: let location, activity and player state steer the soundtrack. Forest exploration, cooking, fighting and remaining HP could each influence the prompt. He also describes Space DJ, where planets represent musical prompts and proximity changes their influence. Nearby Christmas music and Viking metal can produce surprising mixtures. The cookbook’s role is to make such combinations approachable through working examples, not just isolated model calls.

54:0854:25
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

54:08 · section reference included

Availability and safety constrain the usable pipeline

An audience member’s company remains on Nano Banana 1 because it can offer employees only models hosted in Europe. Newer preview models are therefore unusable for that organization despite their capabilities. Vernade attributes the problem to Google Cloud’s global-endpoint policy for preview models. He expects that rule to persist and is advocating faster general availability instead.

Rapid successor releases complicate that route: in his example, Gemini 3.1 arrives before Gemini 3 reaches general availability, effectively restarting the preview progression. Vernade calls the European deployment problem a top advocacy priority, but gives no delivery date or commitment. For a residency-constrained organization, regional availability is a requirement to check before building around a new model.

Another participant has run Frankenstein in a retro-gaming style and obtained game-like characters. Vernade points out that graphic material may be toned down or refused, which helped motivate his choice of children’s books—although earlier restrictions on images of children had complicated those examples too. The background Chatterer run then reveals its futuristic neon squirrel. Style is flexible; source content and applicable restrictions still determine which requests can complete.

58:5259:14
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

58:52 · section reference included

Use lyrics early and images as musical input

The extended Lyria demonstrations expose another useful pipeline boundary. Requesting text and audio returns lyrics before music. In the workshop’s interface, switching from generate_content to generate_content_stream lets the application consume those lyrics while audio is still being generated. A title or cover-image job can start from the lyrics without waiting for the full response. Current Lyria documentation instead illustrates interactions.create and structured response helpers; that is a newer request surface than the one shown in the recording.

The text output includes lyric timing, not just words. In the displayed example, a subsequent line begins at 4.8 seconds, and the final timing leaves 1.2 seconds without words. Such boundaries provide synchronization data for a player while preserving space for the instrumental ending.

Lyria can also take an image as input. Vernade notes that the earlier chapter illustrations could have helped condition their soundtracks, then demonstrates the idea with a photograph of a handwritten pot-au-feu shopping list. A request for an epic quest with opera voices transforms grocery shopping into dramatic stakes: beef shank and ribs become prizes, vegetables enter the story, and wilting celery signals urgency.

Colab displays a photograph of a handwritten pot-au-feu ingredient list beside a pencil, with audio playback controls underneath.Open full source frame
A pot-au-feu ingredient list appears above an active audio player.

The performance uses multiple voices and turns bouquet garni into a recurring chorus. The distinctive result comes from combining concrete image content with a strong musical interpretation: the ingredients supply the subject, while the opera-quest instruction supplies the treatment.

1:03:051:03:20
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

1:03:05 · section reference included

Detailed instructions still need listening tests

Returning to Interactions, Vernade walks through a request with a model, input and response modalities. He objects to renaming the familiar content field, but the operational change is the returned interaction ID that can chain later requests. The output he opens does not visibly expose that ID, so the walkthrough explains the request shape without completing a visible chaining demonstration.

Musical prompting then becomes more specific. Vernade requests tempo control and a Risset accelerando illusion: music that seems to become faster continuously, potentially useful for exercise. He next tries a sequence of fast acoustic guitar for the first ten seconds, piano for the next ten, then full band. On listening, he explicitly says the output is not following the requested sequence. A timed arrangement prompt is a direction to the model, not an enforced timeline.

Section labels offer another way to describe development: intro, verse, outro, chorus and bridge. But the short clip leaves little room for that structure. Vernade moves to a full-song example with a brief intro, specified scale and intensity, and later verses. Playback begins slowly, then introduces drums, bass and more groove. He recommends the full-song model for complex arrangements because the clip model compresses its choices to make a short output interesting.

The remaining examples broaden the output beyond conventional songs. Supplied Nano Banana lyrics become a song about fruit switching into AI mode. A Shakespeare-like performance uses the model’s literary knowledge, although the attempted spoken treatment still contains music; Vernade says stronger instructions can request speech without background music.

A multilingual example switches languages within the same generation. Vernade notes occasional pronunciation errors, then plays a bilingual explanatory song. Across these experiments, the prompt carries the subject, language, structure, delivery and instrumental treatment. The examples are most informative when heard alongside the instructions: they reveal both the breadth of control and the places where requested details disappear.

1:06:371:06:46
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

1:06:37 · section reference included

Navigate the music instead of submitting another song

The session ends by opening the Space DJ application described earlier. Its glowing musical locations and weighted prompt list turn spatial movement into continuous steering. An audience member asks how to try it; Vernade identifies it by name, then navigates toward sea shanties. The model produces vocalizations even though he describes RealTime as not intended for voices.

Space DJ interface with labeled glowing points including sea shanties and British dance, a weighted prompts list, and playback controls.Open full source frame
Space DJ displays music styles as glowing points in a spatial interface.

As the style changes, apparent vocal continuity sometimes survives. Autopilot moves through the musical space without manual navigation, continuing the stream. Vernade explores Southern rock and Australian hip-hop before seeking a faster change. The application makes the ongoing prompt mixture tangible: moving through the interface changes the performance already in progress.

Vernade says the demonstrated session ends after ten minutes; he does not distinguish whether that limit belongs to Space DJ or the underlying API. An application meant to provide continuous music must therefore account for session lifecycle as well as musical steering. He notices the absence of a search button, finds speed metal and stops the performance. The final example turns generative media from a sequence of completed assets into a changing interface—one whose usefulness depends on navigation, continuity and operational limits as much as on the model’s output.

1:13:531:13:56
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

1:13:53 · section reference included

Resources

From the talk

Updates since the talk

Transcript download
Full transcript

Loading transcript…