Veo for ads: plan shots and generated sound

Dialogue inside a generated scene creates a different production problem from silent footage beneath narration. Decide which route the ad needs before writing a Veo brief, then review sound and picture independently. This proposed workflow uses Google’s documentation for version context and controls; it is not a report of our own Veo tests. Compare the intended shot role with the other routes in our model selection guide before generating the complete ad.

Scope the facts to Veo 3.1

Google’s Google: Veo 3.1 API guide describes Veo 3.1 with generated audio, image-led direction, and frame-specific generation. The Google DeepMind: Veo gives the maker’s broader model overview. This guide uses those limited facts, checked on 11 October 2026, without treating every Veo model or every hosting interface as identical.

Confirm the exact model identifier and available options in the interface you intend to use. A product family name is insufficient for a production recipe. If you change versions after a successful test, review the relevant input and output settings again rather than copying the old acceptance decision onto the new request.

Choose between scene dialogue and narration

Generated dialogue ties the spoken performance to the pictured moment. Separate narration keeps the explanation independent of the shot’s mouth movement. Neither is inherently the right choice for every ad. Choose according to what the viewer must understand and what evidence the picture supplies.

For a hypothetical product introduction, a silent room scene under approved narration may be enough. If a visible character must deliver a particular line, evaluate the exact words, pronunciation, and mouth movement as part of the shot. Do not approve a visually appealing scene whose dialogue changes the claim or introduces an invented personal experience.

Source: suggested production checklist, October 2026.

1. Speaks on screen
2. Narrated over

Write dialogue as a bounded performance

Specify who speaks and what the line means before describing atmosphere. Keep the first test simple enough that a reviewer can identify every word. A scene with several speakers, an action demonstration, and a complex camera move tests too many dependencies at once.

Store the approved text separately from the generation prompt. That makes it possible to compare the output’s spoken words with the actual claim rather than with a long descriptive paragraph. If the line changes, treat it as a script revision. The model’s improvised wording is not a reason to relax evidence requirements.

A hypothetical recipe brief can ask for a speaking performance or use silent action beneath separately approved narration. Those routes create different review tasks. A speaking candidate needs its words, pronunciation, mouth, and action checked together; narration allows the editor to approve the voice independently before placing the picture. These are alternatives for a proposed Veo evaluation, not model outputs or evidence that either route has passed a production test.

Use frame controls for a visible purpose

An initial image can establish composition; a desired endpoint can express where the action should finish. Verify support in the selected Veo 3.1 route before depending on those controls. Frame direction is a constraint to test, not a promise that the entire path between frames will be physically convincing.

1. Start frame
2. Shot

Make the endpoints compatible. A closed product on a table and the same product open in the same place may express a useful transition. Changing the room, wardrobe, camera height, and product state simultaneously makes it harder to judge the intervening motion. AI video start and end frames: plan the shot explains the separate problem of endpoint planning.

Review the audio independently

Listen without watching the picture first. Compare the words with the approved line, note any added speech, and ask a fluent reviewer to assess unfamiliar names or languages. A visually convincing mouth does not establish correct pronunciation.

Then watch with sound and check synchronization across the intended usable interval. If dialogue passes but ambience masks a word, revise the audio plan rather than silently rewriting captions. If the voice fails, a separate narration track may preserve a useful silent scene, provided the picture does not visibly contradict that choice. Keep this alternative explicit in the brief.

1. 01Select Veo versionConfirm audio and frame support
2. 02Choose the audio routeDialogue or narration
3. 03Review speech aloneMeaning and pronunciation
4. 04Review the combined shotMotion and synchronization

Review the picture as a second approval

Compare the first, middle, and last usable frames with the intended setting and object. Inspect hands, edges, text areas, and changes in scale. Play the complete interval too, because a contact sheet can miss an abrupt movement between otherwise plausible stills.

Check whether the shot still serves its sentence after trimming. A beautiful establishing view may be too distant to show the product action the script describes. If the demonstration is essential evidence, use real footage or a clearly supported visual. Treat illustrative animation as explanation, not proof that a physical product performed an unfilmed task.

Generated dialogueSeparate narration

• Generated dialogueSpeech belongs to the pictured sceneSeparate narrationVoice remains independent of the shot
• Generated dialogueReview words and mouth movementSeparate narrationReview explanation and visual relevance
• Generated dialogueKeep accepted scene audioSeparate narrationMute unwanted source speech

Keep comparisons fair and narrow

Hailuo for ads: image-to-video shot planning and Lip-sync for dubbed ads: when the mouth must match cover a camera-controlled model route and a separate dubbing decision. To compare them, preserve the production question: the same object, intended action, camera relationship, and acceptance criteria. Adjust syntax to the interface without changing the question midway.

A result can establish that one returned shot fits a specific edit. It cannot establish that one family is superior for all advertising. Record the version, interface, input references, and observed limitations. Rejected audio and accepted visuals should remain separate findings so the next producer knows which part of the route was useful.

1. Setting A
2. Setting B

Package the chosen route for editing

Save the approved text, prompt, references, returned media, selected interval, and audio decision. Identify whether the editor should retain scene sound, mute it, or use a separately approved voice. Put that instruction in the recipe rather than relying on someone to remember a review conversation.

Turntake’s models plan labels the planned Turntake model connection as Coming. This guide does not claim firsthand Veo results or self-service Veo generation inside Turntake. Its deliverable is a bounded test plan that keeps the spoken claim, the pictured action, and the chosen audio route independently reviewable.

Who wrote this We made the ads first, for SmartSolar in Denmark, Sweden and Ireland. Then we built Turntake. Where Turntake comes from