Veo 3 for developers

How do you create an advertisement—or recreate an experience you see every day—when the interface is a description rather than a production timeline? Paige Bailey, introducing herself as engineering lead for Google DeepMind’s DevRel team, approaches that question through the generative media team’s models. Veo 3, Imagen 4 and Lyria 2 cover video with audio, static images and music, respectively.

The broader possibility is to make video a useful medium for expressing ideas, not merely a finished entertainment product. Bailey invokes Andrej Karpathy’s framing of video as a surface for communication, education and human creativity. Before introducing native audiovisual generation, she starts with the creative controls in Veo 2: the tools for deciding what a scene contains and how it moves.

Reference-powered video separates what you want to preserve from what you want to change. Supply a person and an environment as references, then compose the person into that setting. Bailey describes this as a way to produce a stylized, deliberately composed result, and says reference-powered generation was available through APIs and Flow at presentation time.

The little-monster example makes the operation concrete. A purple-and-yellow reference character appears in several environments, with descriptions controlling where it goes. The reference supplies a recurring subject; the prompt supplies a changing setting. The slide places the reference above four results, making character consistency across scenes directly comparable.

Bailey reports that most human raters preferred Veo in some reference-video side-by-side comparisons with Runway Gen-4 and Kling. The presentation does not establish a numerical score or a complete rating protocol here. The useful distinction is that this comparison concerns reference-powered generation: producing an appealing clip while following supplied visual references.

References can also guide style. Upload an image to preserve its visual treatment or compose styles together, then direct the camera through natural language and API controls. Bailey names moving back, moving right, rotating up and zooming in. These are separate creative decisions: a reference constrains appearance, while a camera instruction constrains how the viewer encounters the scene. She also notes a practical developer gap: too few code samples demonstrating the available controls.

Outpainting starts with a limited view and generates what could exist beyond its boundaries. Bailey connects it to the Sphere’s Wizard of Oz project: take a scene or an individual video frame and imagine the surrounding world while keeping the extension visually consistent with the visible portion. The problem is not simply making the picture larger; it is making the newly visible area belong to the same scene.

Object editing changes the contents rather than the boundaries. Bailey shows adding objects and removing objects, including further removal examples, and describes these Veo 2 capabilities as available to test through APIs. Together, outpainting and object editing let a creator revise an existing scene instead of requesting a completely new composition.

Character controls move from scene composition to performance. Reference facial and mouth movements can guide a character’s animation. A script and voice tone can guide lip movements aligned with the sound and appropriate to the location. Bailey then shows input-image motion and design-change examples across Veo 2 and Veo 3: the starting image anchors the appearance while the generated sequence changes its motion or presentation.

First-and-last-frame generation provides a different constraint: specify the beginning and ending images, then have Veo generate the transition between them. After another benchmark slide, Bailey shows endpoint examples stitched into continuous video. These controls address distinct production needs:

The endpoint approach is especially concrete: the destination is already an image, rather than something the model must infer entirely from prose.

Veo 3’s central distinction is native audiovisual generation. Bailey describes the model as composing tokens across video and audio together, rather than calling a separate tool to add sound afterward. She compares this with Gemini’s native audio and broader multimodal capabilities: output need not stop at text or code.

Prompt adherence matters across both modalities. The demonstrations begin with a little llama, then show that generated sound can include music as well as background noise and very subtle effects. The target is therefore more than a moving picture with an arbitrary soundtrack: it is a scene whose visual events and audible details are generated together.

That coherence is difficult even before adding sound. A prompt may contain nuances that generation misses; a character may jump between frames; a wall may disappear and expose a background that should remain hidden. Bailey describes Veo 3 as improving both stylistic consistency and contextual consistency—the persistence of how a scene looks and what exists within it.

She places this progress in a research lineage that includes GQN and WALT. GQN concerns scene representation and rendering, so this is a broad lineage rather than a disclosed Veo architecture. Alongside generation, Bailey describes human-visible watermarks and SynthID watermarks for synthetic images and video. She also emphasizes collaboration with artists, including Darren Aronofsky, alongside musicians working with Lyria and artists working with Imagen.

Imagen 4 addresses the static-image part of the workflow. Bailey’s examples range from humans to whales and puppies, then move from realism to detail across different styles. Typography is part of that detail: the Alamo Square and Mission stamp designs combine an illustrated treatment with readable place names. Bailey imagines the designs as laptop stickers, a small but concrete use for generated imagery beyond a gallery of samples.

The image section closes with designs from a collaboration with Ross Lovegrove. As with the video examples, the emphasis is on creators shaping an output’s appearance, rather than treating image generation as a single undirected request.

Bailey presents Lyria 2 as a high-fidelity music model with granular control over musical inputs, outputs, tones and styles. Music AI Sandbox gives those capabilities a visual workspace. Its editor, with colored waveforms, text fields and a clip list, resembles the kind of music-production environment an Ableton user would recognize.

MusicFX, from Google Labs, offers natural-language composition of beats and supplies the background music for Bailey’s later commercial experiment. Lyria RealTime brings another emphasis: musical interaction developed with musicians, including Jacob Collier and Toro y Moi.

The musician clip that follows argues for learning by playing music and playing with music. Beginners do not have access to the whole musical landscape on day one; exploratory tools can broaden what they can try. The discussion connects music to mathematics, physics, history, geography, the human body and language, treating musical experimentation as a way into a wider set of patterns and relationships.

An audience member asks whether that clip was Veo 3 and comments on visible veins. Bailey initially says it was not, then qualifies that some visuals might have been generated with Veo 3. She does not identify which portions. Returning to Lyria, she emphasizes development with the creative industry and SynthID watermarking for the resulting music assets.

To make progress visible, Bailey holds the requested scene constant: a raccoon wearing a black jacket, dancing in slow motion in front of the pyramids. She starts with WALT, dating the example to 2023 and the underlying research activity to around 2022. Its output is visibly choppy in her account, and she describes the short sequence as not even reaching eight seconds.

The sequence then moves through LTX-Video, which she dates to 2024 and accessed through Hugging Face; Kling 2.0, dated to 2025 and accessed through fal; and Veo 2, dated to 2024. For Veo 2, Bailey likes the raccoon but questions how well it is actually dancing. That distinction matters: a convincing subject is only part of satisfying a prompt that also specifies an action.

Finally, the same prompt goes through Veo 3. The sequence is a qualitative comparison of examples, not a controlled numerical ranking. Its value is the repeated task: the subject, clothing, action, speed and location remain recognizable requirements as the models change.

Image-to-video generation begins with a visual anchor instead of asking language to define everything. Bailey shows a woman with a puppy, then a western-style woman walking slowly forward toward what she describes as a gunfight. Other examples take a stylized person from a single frame and direct actions such as running toward the camera or being lifted into the sky by a tractor beam. Natural language specifies what should happen to the supplied subject.

Prompt rewriting works on the language side of that interface. The short raccoon sentence is expanded into a fuller description that Veo can more readily use. This adds a preparation step between the user’s concise intent and the generation request: elaborate the description, then generate from that richer specification.

Bailey asks the audience to watch the resulting example with music, sound effects and background noise in mind. She highlights the visual detail, including apparent reflections in the subjects’ eyes. The demonstration brings the two threads together: more fully specified visual direction and sound generated alongside the image sequence.

The Paris team’s sequence broadens the demonstrations to music and dialogue. Bailey compares the musical style to Daft Punk, then points out a haircut floating at the top of a head. The examples can be expressive while still containing conspicuous visual oddities.

Her explanation of the dialogue supplies the jokes for viewers who do not speak French or Spanish: Spanish veo means I see, while another character responds that he does not understand what is being said. The sequence includes repeated references to artificial intelligence and an interview, then an alien exchange about whether to contact humans who are about to create AGI. The response is that the humans are not ready yet. These are demonstrations of generated audiovisual scenes carrying timing, language and humor, not just recognizable objects.

The access discussion describes the presentation’s rollout state, rather than a current availability matrix:

• Google AI Ultra: access in multiple countries, with the UK described as recently added.
• Google AI Pro: a limited number of uses through the Gemini mobile app.
• Vertex AI: Veo 3 in private preview.
• Veo 2: available through Vertex AI or the Gemini APIs.

Bailey expresses hope for AI Studio access and offers an early-access form through a QR code. That hope should be read in the Veo 3 context: the earlier Veo 2 developer release had already announced Veo 2 in AI Studio.

The code walkthrough exposes a small set of production choices: which model to call, where to deposit the generated video, whether to start from an input image, and which aspect ratio to request. An output bucket is the destination; a starting image is a generation input. Keeping those roles separate makes the request easy to reason about.

A Python request in the historical Gemini API style illustrates the model-and-configuration boundary. Here the repeated raccoon prompt is the text input, and aspect_ratio controls the output shape:

from google import genai
from google.genai import types

client = genai.Client()

operation = client.models.generate_videos(
    model="veo-2.0-generate-001",
    prompt=(
        "A raccoon wearing a black jacket dancing in slow motion "
        "in front of the pyramids."
    ),
    config=types.GenerateVideosConfig(
        aspect_ratio="16:9",
        number_of_videos=1,
    ),
)

print(operation.name)

This submits a generation operation; it does not make a completed video immediately available. The historical Veo 2 example illustrates the small request surface Bailey emphasizes, while her onstage walkthrough also discusses output-bucket and optional-image configuration.

The final experiment returns to advertising. Bailey had intended a live demo, but time and uncertainty about the demo lead to a walkthrough. The source is a Chick-fil-A commercial whose speaker also happens to be named Paige—Bailey explicitly says it is not her. The copy praises crispy breading, a tender filet, warmth and satisfaction. Recreating it requires more than matching a sandwich image: it requires shots, speech, music and a finished sequence.

The Veo 2 workflow begins by turning that finished source into a production plan:

1. Give the original video to Gemini and ask for a detailed reconstruction plan.
2. Segment the plan into generation prompts that fit the workflow’s eight-second Veo 2 clip limit.
3. Use MusicFX to create a background track combining Down Home Farm and Slow Guitar.
4. Bring the generated material into Camtasia and stitch it together with transitions.

Bailey describes the process as relatively quick, but still requiring substantial thought and work to reach final assembly.

She plays the assembled version, including the recreated introduction and sandwich description, then identifies Gemini text-to-speech as the speech source. The complete production therefore combines Veo 2 video, separately generated music, separately generated speech and manual editing. Bailey prefers her assembled result to the original commercial, but the operational point is how many outputs she had to coordinate to get there.

With Veo 3, Bailey describes a shorter path: take the original video, generate a description, and submit that description to Veo 3 in Flow. The shown output includes the Paige introduction and the copy about crispy breading and a tender filet. Bailey attributes the shown Veo 3 commercial result to one prompt submitted in Flow. Her account describes the reduction in manual work; it does not provide measured generation latency or repeatability.

The repeated playback makes the practical consequence of native audio concrete. In this example, the video and spoken performance arrive together, removing the separate speech-and-video assembly step that occupied the Veo 2 workflow. The result is still judged by how well it recreates the intended scene and delivery, but the creator has fewer independently generated pieces to coordinate.

Bailey closes with a commitment to broaden access and add more steerability controls, without a delivery schedule. The farewell provides one final demonstration: the Paige waving on screen is not footage of her waving. It is a static photograph animated with Veo 3—the same image-to-video mechanism now applied to the presenter herself.