# AI video prompt generator

URL: https://dreambench.ai/tools/ai-video-prompt-generator

A good video prompt reads like a shot list: who is in frame, what they do, where the light comes from, how the camera moves, what is heard. This page explains each part and the mistakes that blur a clip; its generator writes the prompt in that order, for any model.

## What is an AI video prompt generator?

It turns your choices for the subject, action, light, camera and sound into a prompt for any video model, in the order models read best, with presets for Veo 3.1, Seedance 2.5 and Kling 3.0. It is part of every DreamBench plan, from $9.99 a month.

## How to write a video prompt

1. Describe the shot: Who or what is in frame, what they do, and where: one subject, one action, one place.
2. Direct it: Pick the light and its direction, the shot size and camera move, the look, the mood and the sound.
3. Copy or run it: The prompt is assembled in the order video models read best. The first time, create a free account and choose a plan, then copy it for any tool, or run it on DreamBench.

## How a video prompt is built

1. Subject and action first: Who or what is in frame and the one thing they do, in plain words: a flower seller wraps a bunch of tulips. A model builds everything else around that clause.
2. The setting in one clause: Where and when: a street market on a wet evening. One place and one stretch of time; a new place is a new clip.
3. The light and where it comes from: Name the source and its side, and what falls to shadow. Left unsaid, models light a scene flat from the front.
4. One shot size and one camera move: Wide, medium or close-up, a lens, and a single move: a slow push-in, a lateral track. Two moves smear into neither.
5. The look, once: Photoreal on film, documentary, anime, stop-motion or black-and-white: one register, said once.
6. Behavior, not mood: A feeling becomes something a camera can see: she grips the rail, rather than she is nervous.
7. What is heard: Ambience, one spoken line in quotes, a music bed or silence, where the model can leave sound out, landing in step with the action.
8. The length and the frame: Written for any model, the prompt closes with them; each model's preset opens with them, as that model's own examples do.

## Example prompts

**One shot for any model.** This page's own order, for any model: scene, light, camera, look and sound, then the length and the frame.

```text
An old flower seller in a waxed jacket wraps a bunch of tulips in brown paper and hands it across the stall, at a street market on a wet evening. Light: night, lit only by the practical lights in the scene from camera-left, so the subject's camera-right side falls to shadow; the exposure holds the subject; no fill from the camera side. Medium shot on a 50mm lens at a couple of metres: natural proportions, the background softening behind the subject. One move only: a slow push-in that ends close on the subject. Photoreal, shot on film: natural colour, gentle grain, true skin. Sound: ambience and one spoken line, "Last bunch of the day. Mind the rain." An 8-second clip, 16:9. No text, no captions, no watermark.
```

**The frame first, the sound after (Veo 3.1).** The frame first, then what is heard, because Veo makes sound with every clip.

```text
Widescreen 16:9, one 8-second shot. An old flower seller in a waxed jacket wraps a bunch of tulips in brown paper and hands it across the stall, at a street market on a wet evening, and says, "Last bunch of the day. Mind the rain." Sound: the line clear over the scene's own ambience, every sound landing in step with the action; no music. Light: night, lit only by the practical lights in the scene from camera-left, so the subject's camera-right side falls to shadow; the exposure holds the subject; no fill from the camera side. Medium shot on a 50mm lens at a couple of metres: natural proportions, the background softening behind the subject. One move only: a slow push-in that ends close on the subject. Photoreal, shot on film: natural colour, gentle grain, true skin. No text, no captions, no watermark.
```

**One 15-second take (Seedance 2.5).** Fifteen seconds in one take, with room for the whole exchange; the sound comes last.

```text
Widescreen 16:9, one continuous 15-second take with no cuts. An old flower seller in a waxed jacket wraps a bunch of tulips in brown paper and hands it across the stall, at a street market on a wet evening. Light: night, lit only by the practical lights in the scene from camera-left, so the subject's camera-right side falls to shadow; the exposure holds the subject; no fill from the camera side. Medium shot on a 50mm lens at a couple of metres: natural proportions, the background softening behind the subject. One move only: a slow push-in that ends close on the subject. Photoreal, shot on film: natural colour, gentle grain, true skin. Sound: ambience and one spoken line, "Last bunch of the day. Mind the rain." No text, no captions, no watermark.
```

**The line in its own language (Kling 3.0).** One take, the line with its language, and the subject held steady through the move.

```text
Widescreen 16:9, one continuous 10-second take with no cuts. An old flower seller in a waxed jacket wraps a bunch of tulips in brown paper and hands it across the stall, at a street market on a wet evening, and says in English, "Last bunch of the day. Mind the rain." Medium shot on a 50mm lens at a couple of metres: natural proportions, the background softening behind the subject. One move only: a slow push-in that ends close on the subject. The subject and everything around it keep their shape and look as the camera moves. Light: night, lit only by the practical lights in the scene from camera-left, so the subject's camera-right side falls to shadow; the exposure holds the subject; no fill from the camera side. Photoreal, shot on film: natural colour, gentle grain, true skin. Sound: the line in English, the lips in time with every word, over the scene's own ambience. No text, no captions, no watermark.
```

**A square product close-up.** Five square seconds for a product page: an extreme close-up, studio light from above, one push-in and a music bed.

```text
A glass bottle of olive oil turns slowly on a stone plinth as a single drop runs down its side, on a dark kitchen counter. Light: a single soft studio key light from above, short shadows falling straight down beneath every form; the exposure holds the subject; no fill from the camera side. Extreme close-up on an 85mm lens from six metres: the background compressed and soft, only the subject sharp. One move only: a slow push-in that ends close on the subject. Photoreal, shot on film: natural colour, gentle grain, true skin. Sound: a quiet music bed under the ambience, no dialogue. A 5-second clip, 1:1. No text, no captions, no watermark.
```

**A handheld street moment.** Ten vertical seconds filmed on the street: overcast daylight, a handheld follow, the documentary look and the street's own sound.

```text
A cyclist in a red jacket weaves through slow traffic and glances back once, on a grey city street after rain. Light: soft overcast daylight from above, short shadows falling straight down beneath every form; the exposure holds the subject; no fill from the camera side. Medium shot on a 35mm lens at a few metres: the subject and the room both readable, little distortion. Handheld, following the subject: the operator's breath and small corrections, never random shake. Documentary register: available light, unposed, nothing staged for the lens. Sound: the scene's own ambience, no music. A 10-second clip, 9:16. No text, no captions, no watermark.
```

## Common mistakes

- Adjectives instead of directions: Beautiful, cinematic and high quality tell a model nothing it can film. Name the light, the lens and the move.
- Two camera moves in one clip: A push-in that turns into an orbit smears into neither. One move per clip.
- A whole story in one clip: Three places and a jump in time give one muddled clip. One place and one stretch of time per clip; the composer writes one prompt per scene.
- Moods a camera can't see: Sad or nervous are invisible. Write what the body does: shoulders drop, a hand tightens.
- Light left unsaid: With no light named, models light everything flat from the front. Name the source, its side and what falls to shadow.
- Asking for what the model can't make: A 20-second Veo clip or a silent one: each model has its own lengths, frames and sound, and the presets offer only those.

## Works with

- What it takes: Your own words: who is in the shot, what they do and where. No photo is needed.
- What you get: A prompt as plain text.
- DreamBench's composer: Paste it into Video, which sends it as written, or make it a film’s brief.
- Google Flow, Kling and Dreamina: Each takes the same plain-text prompt.
- A prompt library or a script: The parts come in a fixed order, so a series of prompts stays consistent.

## Which model it writes for, and its limits

The AI video prompt generator writes for any video model; its presets write for [Veo 3.1](https://dreambench.ai/models/veo-3.1), [Seedance 2.5](https://dreambench.ai/models/seedance-2.5) and [Kling 3.0](https://dreambench.ai/models/kling-3.0), within what each can make on DreamBench.

| | Veo 3.1 | Seedance 2.5 | Kling 3.0 |
|---|---|---|---|
| Lengths | 4, 6 or 8 seconds | 5, 10, 15, 20 or 30 seconds | 3, 5, 10 or 15 seconds |
| Frames | 16:9 or 9:16 | 16:9, 9:16, 1:1, 4:3 or 21:9 | 16:9, 9:16, 1:1, 4:3 or 21:9 |
| Sound | Always on: ambience, or ambience and a spoken line | On or off: ambience, a spoken line or a music bed | On or off: ambience or a spoken line |
| Shots | One shot, no cuts | One take, or two or three linked shots from 10 seconds | One continuous take, no cuts |
| Dialogue | One line, said in the scene | One line, said in the scene | A line lip-synced in English, Chinese, Japanese, Korean or Spanish |
| Prompt format | Plain text or JSON | Plain text | Plain text |

- Written like a shot list: Subject and action first, then the setting, the light, the camera, the look and the sound: the order models follow most reliably.
- Light as a lock: The generator names the key light and its direction, against the flat, evenly lit image a model adds by default.
- One move at a time: Stacked camera moves smear: you pick one move and one shot size, and the prompt says what the optics do.
- Behavior, not adjectives: A mood becomes something the body does on camera; a model follows a visible outcome better than a feeling.

**Where it fails.**
- A prompt sets what a model aims for, not what it makes: when a detail is ignored, run it again or say it more plainly.
- Several people interacting is the least stable scene for video models; one clear subject holds best.
- ByteDance says complex motion in Seedance 2.5 can look physically off, and Google that Veo 3.1's speech in short lines is still being improved.

## AI video prompt generator FAQ

- **How do I write a prompt for AI video?** To write a prompt for AI video, say who is in frame and the one thing they do, then where and when, the light and its direction, one shot size and one camera move, the look once, the mood as behavior, and what is heard. Close with the length and the frame. Models follow that order best and ignore vague adjectives.
- **Does it make the video, or only the prompt?** The AI video prompt generator makes only the prompt, not the video: it calls no video model. Copy the prompt into any tool, or run it on DreamBench, where the composer makes the clip on Veo 3.1, Seedance 2.5 or Kling 3.0 and shows its price before it makes anything.
- **Does the prompt work with Runway, Luma and other models?** Yes. The any-model prompt is plain text in the order every current video model reads, so it pastes into Runway, Luma, Hailuo or any tool with a prompt box. The presets write for Veo 3.1, Seedance 2.5 and Kling 3.0, within what each makes on DreamBench; for another model, use the any-model prompt and check that model's lengths and frames yourself.
- **How long should a video prompt be?** Long enough to direct every element the clip needs and no longer: for one clip of five to ten seconds, 60 to 150 words covers subject, setting, light, camera and sound without repeating itself.
- **Should I use a negative prompt?** Mostly, no: describe what you want instead. Google's Veo guide says to state exclusions positively, a desolate landscape with no buildings rather than no man-made structures. This generator closes every prompt with one short line, no text, no captions, no watermark, and says everything else positively.
- **Can I describe several scenes in one prompt?** Not with this generator: it keeps a clip to one place and one stretch of time. Seedance 2.5 can cut between linked shots, and Google's Veo guide shows timed shots inside one clip. DreamBench's composer makes multi-scene videos by writing one prompt per scene.
- **Can it write a prompt from an image or a video?** Not from an upload: it writes from your choices and doesn't look at a picture or watch a clip. When a picture is the first frame, it already sets the subject, the light and the look, so describe only what moves and what the camera does; to match a clip you have, describe what you see. The AI image to video tool takes such a note with your photo.
- **Can ChatGPT or Gemini write a video prompt instead?** Yes, and without a plan here, if you tell them what the model needs: an assistant writes fluent prompts but doesn't know each model's lengths, frames and sound on DreamBench unless you say. This generator needs a DreamBench plan, shows the price in credits first, and writes only what the model can make.
- **How much does the AI video prompt generator cost?** Each prompt uses 11 credits from your DreamBench plan, and the same prompt again uses none. Starter, $9.99 a month, includes 1,100 credits a month, about 100 prompts, plus AI video and the image tools.
- **Do I need an account to use the AI video prompt generator?** Yes: a free account and a DreamBench plan. The tools are part of every plan, from Starter at $9.99 a month.

Calls no model; 11 credits a prompt. DreamBench is an independent platform; its tools come with every plan (https://dreambench.ai/pricing). Copy last checked 2026-10-14.