Veo 3 introduced short generated scenes with their own sound. For a creator, that means briefing the picture and audio together, then checking both. The original 3.0 API endpoints have since retired, so a usable workflow now also requires identifying which version your chosen product exposes.
An eight-second bakery scene shows how to plan that combined picture-and-sound job below. First, separate the original model specifications from later capabilities so you do not choose a workflow that depends on an unavailable endpoint.
Veo 3 specifications and what changed in 3.1
Google's model lifecycle table records 30 June 2026 as the retirement date for the original veo-3.0-generate-001 and Fast endpoints, with Veo 3.1 GA as their replacement. These are cloud endpoint dates, not a claim about every third-party product using the Veo name.
| Property | Original Veo 3 cloud model | Veo 3.1 GA cloud model |
|---|---|---|
| Model identifier | veo-3.0-generate-001, retired | veo-3.1-generate-001 |
| Inputs | Text prompts; Google's model card also describes image input | Text, image, first/last-frame and asset-reference workflows |
| Generated clip duration | 4, 6 or 8 seconds | 4, 6 or 8 seconds; reference-image workflow is 8 seconds |
| Output resolution | 720p or 1080p | 720p, 1080p or 4K in this standard GA model |
| Shape and frame rate | 9:16 or 16:9; 24 fps | 9:16 or 16:9; 24 fps |
| Sound | Generated audio with video | Generated audio with video |
| Continuation and asset references | Not supported by the original cloud model | Supported, subject to the particular workflow's input rules |
The original specifications are retained in Google's Veo 3 model record; input and output types are also described in the Veo model card. The newer column follows the Veo 3.1 GA specification checked on 31 October 2026. Do not transfer its 4K and continuation options to an old Veo 3 example or assume that Fast and standard variants share every output option.
For a vertical Reel, 9:16 controls the picture shape; 1080p describes the resolution option; 24 fps describes the output cadence. None gives the model a guarantee of exact wording, character identity or physically correct contact. Google's model card identifies consistency during complex scenes and motion as a remaining limitation. The practical response is a manageable action and a clear review of its difficult moment.
Check access before adapting an old example
Google's current video-generation documentation says the Veo 3.1 Preview endpoints in the Gemini API shut down on 22 October 2026. It directs developers to Gemini Omni Flash or the general-availability models through its cloud platform. That notice concerns those preview endpoints; it does not establish that every product carrying the Veo name stopped working.
The cloud text-to-video guide lists supported model identifiers. Check the selected model, region and product before copying a request. A sample labelled Veo 3.1 should not be silently relabelled as a demonstration of the original Veo 3.
| What you see | What to verify | Why it changes the job |
|---|---|---|
| A consumer video interface | Its visible model, available inputs and current account allowance | The product may expose only some model capabilities |
| An API example | Model ID, endpoint status, billing and accepted input schema | An old preview example may no longer run |
| A third-party Veo integration | Version, export conditions and its own terms | Access through another seller is not identical to direct access |
Do not buy credits because a sample mentions a feature somewhere in the Veo family. Confirm that the feature exists in the exact route you intend to use. Download accepted files and keep the prompt and version with them, rather than treating a generation history as permanent storage.
Choose an input that removes the right uncertainty
A text prompt describes the subject, setting, action, camera and sound. It gives the model room to invent appearance. An image input is useful when the opening composition or a particular prop matters, but it still does not promise exact preservation throughout motion.
For the bakery scene, the defining object is a round loaf with one diagonal score on a wooden counter. A still image can establish that loaf and counter. A text-only route can be enough if you mainly need the action of placing a loaf down and do not need it to match a real product photograph.
Do not ask an eight-second shot to establish a location, introduce three people, show kneading, baking and serving, then deliver a spoken sales message. Split the story into shots or choose one observable action. The model has less ambiguity to resolve when the action fits the time.
Brief one eight-second bakery shot
The scene shows a baker's hands lowering a fresh loaf onto a counter. The camera holds a close view. A soft contact sound and a short line of speech complete the moment. There is no need for a complicated camera path or a second location.
Close view of a round scored loaf held in flour-dusted hands above a worn wooden bakery counter. In one continuous action, the hands lower the loaf onto the counter and withdraw. The loaf keeps its round shape and single diagonal score. The camera stays still at counter height. Soft morning light from the left. Audio: a quiet room, gentle bread contact with wood, then an off-camera adult voice says, “Ready for the first slice.” No music and no on-screen text.
The audio cue is separate from the visual action so it is easy to revise. The line is short enough to leave room for the placement. The prompt avoids asking the video model to render an exact graphic title; add any required title in an editor where spelling and timing remain under your control.
| Time in the edit plan | Visible action | Review question |
|---|---|---|
| 0–2 seconds | Hands hold the loaf above the counter | Are the loaf and hands stable? |
| 2–5 seconds | Loaf lowers and makes contact | Does it meet the counter without passing through it? |
| 5–8 seconds | Hands withdraw; loaf remains | Does the object keep its score and shape while the line finishes? |
These timings guide selection and editing. They are not an assertion that a prose prompt gives frame-accurate control. If the result spends too long on the first action, simplify the brief or choose a shorter movement before demanding more events.
Review picture and audio separately
First watch normally. The scene should read as one clear action. Then mute it and inspect contact, finger shapes, loaf identity and lighting. Check the beginning, the moment the loaf touches the counter and the final frame.
Listen without watching. Does the spoken line say the intended words? Does an unwanted voice or music enter? A visually strong take may still need replacement audio. Keep a clean source and rights-cleared narration available when exact pronunciation or delivery matters.
For a real product, compare the generated appearance against the approved reference. A loaf can tolerate some natural variation; a logo or package label may not. Use authorised graphics in post-production for exact text rather than accepting a near match that changes the claim.
Revise one cause at a time
If the loaf changes shape, reduce simultaneous motion and emphasise the distinguishing score. If the hands obscure the subject, adjust the starting composition. If the speech overruns the action, shorten the line or record it separately. Changing camera, subject, lighting and dialogue together makes it harder to tell which revision helped.
Keep a record with model/version, input image, prompt, available settings and the accepted output filename. A saved seed, when supported, does not establish identical output across every version or service. The seed guide explains that boundary.
Turn the shot into an actual short
Place the accepted shot in the intended vertical or landscape timeline. Add the required title as selectable editor text, balance the final audio and trim the end after the action resolves. Watch the exported file, including its first frame and final audio tail.
For a longer story, use separate shots with explicit continuity notes rather than assuming a single generation will become a complete film. The cinematic planning guide covers that assembly. If your central need is a recurring narrated faceless series, compare it with the ShortsFast production workflow. Shot generation and recurring episode production involve different controls and review work.

