Veo 3.1 prompt guide

The Veo 3.1 prompt generator writes a prompt in this order; this guide explains each part. For the model, see the Veo 3.1 page.

Google’s prompting guide for Veo 3.1 gives five parts. The camera comes first because it decides what the rest must fill; the sound follows in sentences of its own.

1. 1Open with the camera: shot size, angle, one move and, if it matters, the lens.
2. 2Name the subject with the details that must not change.
3. 3Write one action in the present tense, from start to end.
4. 4Add where and when, the weather and the background.
5. 5Set the look, the light and its direction, and the mood.
6. 6End with the sound: any line in quotes after its speaker, then “SFX:” and “Ambient noise:”.

Google DeepMind’s guide for Veo 3 uses the same parts and shows each prompt beside its video, from one line to a long play-by-play. A prompt written for the older model works on 3.1.

Veo reads film language, so name the shot as a camera operator would. These terms come from Google’s prompt guide.

One move per clip: a dolly that becomes an arc in 8 seconds smears both. For dialogue, a static shot or a slow dolly in keeps the face steady.

Every Veo 3.1 clip has sound, made with the picture. Google asks you to describe it in sentences of its own, and its Veo 3.1 guide labels each layer. See native audio.

1. 1Name the speaker, then the line in quotes: A woman says, “We have to leave now.”
2. 2Give a tone where it matters: he says in a weary voice, “Not again.”
3. 3Put each effect after “SFX:”: SFX: the door slams.
4. 4Put the background after “Ambient noise:”: Ambient noise: the hum of a starship bridge.
5. 5Write music as a sound: a gentle orchestral score begins to play.

• One short line in 4 seconds, two in 8: speech runs at two to three words a second.
• Give every line a speaker in shot, or it can become a voice-over.
• Write lines in English: Google’s model reference lists English as the prompt language. Veo moves the lips in the same generation; see lip sync.

Subtitles and garbled speech

Google lists natural speech in short lines as a limit it is still working on. MIT Technology Review found garbled subtitles on clips with dialogue even when the prompt asked for none; Google’s advice was to run it again. So describe a clean frame, leave signs and captions out unless you want words on them, keep lines plain, and listen to every take.

The length is set before Veo reads the prompt, so write only what the seconds hold: a longer action is rushed, a longer line cut off.

Google’s guide also splits one 8-second clip into timed shots, each opening with its range, such as [00:00-00:02]. Two seconds is short, so keep it for when the cuts matter. Veo makes 16:9 or 9:16 only; for square or 21:9, see the Kling 3.0 prompt guide.

Google’s tools take pictures two ways: a first frame to animate, or up to three reference images of one person, character or product to keep in a new shot. See image to video and reference to video.

Google’s reference example names its subject plainly: “A person walks in carrying a vase full of flowers and places the vase on a kitchen table.” References work on Veo 3.1 and Fast, not Lite, and the full model then makes 8 seconds.

On DreamBench your pictures become sheets. Veo 3.1 takes up to 3 with each clip, which then runs 8 seconds; on Fast and Lite, DreamBench draws an opening frame from them.

Google offers first and last frame on all three tiers: a start and an end picture, and Veo makes the move between them, with sound, at 4, 6 or 8 seconds (Google Cloud). Describe the change, as Google’s example does: “a hand reaches in and places a glass of milk next to the plate of cookies”. Keep the frames close in light and framing. See first and last frame.

DreamBench doesn’t offer a start or end frame for Veo: its composer draws each clip from your references, so the prompt’s last sentence says how the shot ends.

Google’s API has a negative prompt field for what you don’t want. Its prompt guide says to list the things, “wall, frame”, never “no walls”; in the main prompt, its Veo 3.1 guide says to describe the result instead.

DreamBench sends Veo no negative prompt field. The prompts it writes carry a short “Negative:” line naming burned-in subtitles where a clip speaks. On the Video page, describe the result and keep any exclusion to one closing line, as the prompt generator does.

Two Veo clips agree only on what both were given. Google’s Veo 3.1 guide makes reference images of the cast and setting first. In the prompt:

• Describe each character in the same words every time: “a tall woman in her sixties with cropped silver hair and a green waxed jacket”.
• Keep the clothes, props, time of day, light and voice.

On DreamBench each character, place and prop is drawn once as a sheet, and every shot it is in starts from it. See character consistency.

Google shows two ways past one shot: timed shots inside one 8-second clip, or separate clips made from the same references. DreamBench takes the second: a film is several Veo generations, one shot each, cut together, so it runs past 8 seconds and any shot can be redone.

• One action per shot, at 4, 6 or 8 seconds; on DreamBench, Veo 3.1 clips that carry sheets run 8.
• Change the framing at each cut, keep a line inside one shot, and let an ambience run across the join.

Write each shot with the formula, or describe the film on DreamBench, which writes, renders and cuts every shot. See multi-shot video.

Paste these into DreamBench’s Video page or any Veo 3.1 tool.

A hook to camera

Vertical 9:16, 6 seconds. Handheld selfie shot. A young man in a gray beanie on a windy pier at dawn looks into the lens and says, “I tried every coffee in town. One won.” Soft cold light. Ambient noise: wind, gulls.

Two people talking

Widescreen 16:9, 8 seconds. Static two-shot. In a bike workshop, a customer in a wet jacket asks, “Ready by Saturday?” The mechanic spins the wheel and says, “Friday, if the parts turn up.” Warm bulb light. SFX: the freewheel ticking. Ambient noise: rain on a tin roof.

Product close-up: a watch

Widescreen 16:9, 6 seconds. Macro lens, slow dolly along a steel watch with a green dial on dark slate. A drop of water runs down the glass as the second hand ticks past twelve. Hard light from the left. SFX: a crisp tick.

Food: garlic in a wok

Vertical 9:16, 4 seconds. Static overhead shot. Sliced garlic drops into hot oil in a black wok, foams and turns gold; steam rises through window light. SFX: a loud sizzle.

An establishing shot

Widescreen 16:9, 8 seconds. Aerial drone shot low over a fishing village on a fjord at first light, rising to reveal snow-capped peaks. Red wooden houses, mist on the water. Photoreal, cool light warming at the peaks. Ambient noise: wind, distant engines.

An animal in action

Widescreen 16:9, 6 seconds. Telephoto lens, low angle. A brown hare sprints across a frosted field, swerves and stops dead, ears up. Low winter sun behind it. SFX: thudding paws. Ambient noise: a crow far off.

Stop motion with music

Widescreen 16:9, 8 seconds. Stop-motion claymation, static wide shot. A clay fox in a knitted scarf builds a snowman; its head falls off and the fox sighs. Warm studio light. A playful plucked-string tune begins to play. SFX: a soft plop.

Four shots in one clip

Widescreen 16:9, 8 seconds. [00:00-00:02] Close-up of a hand pressing a stopwatch. [00:02-00:04] Low tracking shot of spikes striking the track. SFX: spikes on rubber. [00:04-00:06] Medium shot as she leans into the line. [00:06-00:08] Wide high angle: she slows, hands on knees, the stadium empty. Ambient noise: her breath.

From a first frame

The woman in the picture looks up from her book, sees rain starting on the window and smiles. Slow dolly in. SFX: the first drops on the glass. Ambient noise: a quiet café.

Between a first and a last frame

Widescreen 16:9, 8 seconds. Static shot. A potter’s hands draw wet clay up into a tall vase as the wheel spins, then let go. Soft daylight from a high window. SFX: the hum of the wheel.

A two-shot sequence: the gardener

Shot 1. Widescreen 16:9, 8 seconds. Wide shot, slow pan right. A tall woman in her sixties with cropped silver hair and a green waxed jacket walks between frosted cabbages at dawn, a basket on her arm. Cold blue light. Ambient noise: boots on frozen ground, a robin.

Shot 2. Widescreen 16:9, 6 seconds. Static close-up. The tall woman in her sixties with cropped silver hair and a green waxed jacket kneels, brushes frost off a cabbage and says softly, “Just in time.” Dawn turning gold. SFX: a knife through a stalk.

Both shots describe her in the same words and keep the dawn light; the framing changes at the cut.

A weak prompt is usually short: every part left out is a choice Veo makes for you.

• One prompt run twice gives two clips.
• Short lines can sound unnatural, and unwanted subtitles can appear: check every take.
• One generation stops at 8 seconds, in 16:9 or 9:16.
• A real person or product can’t be known from words; it takes a picture you have the right to use.
• On DreamBench’s Video page, a prompt past 2,600 characters is refused.

• One clip: paste your prompt into the Video page and pick Veo 3.1, Fast or Lite, 4, 6 or 8 seconds, 16:9 or 9:16. It goes to Veo as written.
• One prompt: the Veo 3.1 prompt generator writes it in this guide’s order, as text or JSON.
• A film: describe it, add up to 8 reference pictures and pick Veo 3.1. DreamBench writes and renders every shot and cuts them together; redo any shot with a note.
• Not offered: a start or end frame, a negative prompt field, or extending a clip.

The full Veo 3.1 model needs Plus or above; Fast and Lite are on every plan. The price in credits shows before you generate; the first time, create a free account and choose a plan. For prompts from words alone, see text to video.