Google Veo 3 on Replicate: Running the Model via API

行业资讯veo3

Google Veo 3 on Replicate: Running the Model via API

Veo 3 is Google DeepMind's newest text-to-video model, and it stretches what natural-language prompts can produce. With native audio generation, tighter prompt adherence, and striking realism, it reshapes how multimedia content gets made. This guide walks through calling it on Replicate from Node.js, Python, or plain HTTP.

Running Veo 3 in Node.js

The fastest start is a single command:

npx create-replicate --model=google/veo-3

Or build a project from scratch. First install the client library:

npm install replicate

Set your API token (find it in your account settings):

export REPLICATE_API_TOKEN=<paste-your-token-here>

Import and configure the client:

import Replicate from "replicate";
import fs from "node:fs";

const replicate = new Replicate({
  auth: process.env.REPLICATE_API_TOKEN,
});

Then run the model:

const input = {
  prompt: "A breaking news ident, followed by a TV news presenter excitedly telling us: We interrupt this programme to bring you some breaking news... Veo 3 is now live on Replicate. Then she shouts: Let's go!\n\nThe TV presenter is an epic and cool punk with pink and green hair and a t-shirt that says \"Veo 3 on Replicate\"",
  duration: 8,
  resolution: "1080p",
  aspect_ratio: "16:9",
  generate_audio: true
};

const output = await replicate.run("google/veo-3", { input });

// To access the file URL:
console.log(output.url()); //=> "http://example.com"

// To write the file to disk:
fs.writeFile("my-image.png", output);

Running Veo 3 in Python

Install the Python client:

pip install replicate

Set the token:

export REPLICATE_API_TOKEN=<paste-your-token-here>

Import and run:

import replicate

output = replicate.run(
    "google/veo-3",
    input={
        "prompt": "A breaking news ident, followed by a TV news presenter excitedly telling us: We interrupt this programme to bring you some breaking news... Veo 3 is now live on Replicate. Then she shouts: Let's go!\n\nThe TV presenter is an epic and cool punk with pink and green hair and a t-shirt that says \"Veo 3 on Replicate\"",
        "duration": 8,
        "resolution": "1080p",
        "aspect_ratio": "16:9",
        "generate_audio": True
    }
)

# To access the file URL:
print(output.url)
#=> "http://example.com"

# To write the file to disk:
with open("my-image.png", "wb") as file:
    file.write(output.read())

Running Veo 3 over HTTP

If you'd rather hit the API directly, set the token and POST to the predictions endpoint:

curl -s -X POST \
  -H "Authorization: Bearer $REPLICATE_API_TOKEN" \
  -H "Content-Type: application/json" \
  -H "Prefer: wait" \
  -d $'{
    "input": {
      "prompt": "A breaking news ident, followed by a TV news presenter excitedly telling us: We interrupt this programme to bring you some breaking news... Veo 3 is now live on Replicate. Then she shouts: Let\'s go!\\n\\nThe TV presenter is an epic and cool punk with pink and green hair and a t-shirt that says \\"Veo 3 on Replicate\\"",
      "duration": 8,
      "resolution": "1080p",
      "aspect_ratio": "16:9",
      "generate_audio": true
    }
  }' \
  https://api.replicate.com/v1/models/google/veo-3/predictions

Example Output

A run from the prompt above completed like this:

{
  "completed_at": "2025-06-05T23:37:13.835612Z",
  "created_at": "2025-06-05T23:34:55.848000Z",
  "data_removed": false,
  "error": null,
  "id": "wz7b7p4rd1rme0cq89s9c7zj34",
  "logs": "Using seed: 2035471022\nStarting video generation...\nStill generating...\nStill generating...\nStill generating...\nGenerated video in 136.64 seconds\nDownloading video...\nDownloaded video in 0.08 seconds",
  "metrics": {
    "predict_time": 137.607550372,
    "total_time": 137.987612
  },
  "output": "https://replicate.delivery/.../tmpny051d_c.mp4",
  "started_at": "2025-06-05T23:34:56.228062Z",
  "status": "succeeded"
}

Total time end to end: roughly 2 minutes 18 seconds, with the generation itself taking about 137 seconds.

More Examples

Pricing

Replicate bills Veo 3 per second of output, split by whether audio is generated:

Variant Price Equivalent
With audio $0.40 / second 25 seconds for $10
Without audio $0.20 / second 50 seconds for $10

About Veo 3

Veo 3 is Google DeepMind's latest advance in text-to-video generation, with native audio, improved prompt adherence, and realism that the team positions as a redefinition of multimedia content creation.

Key features:

  • Text to image and video — high-fidelity, cinematic visuals straight from text prompts.
  • Native audio generation — ambient noise, sound effects, and dialogue that sync to the visuals with no post-production.
  • Dialogue and lip-sync — characters speak your script with accurate lip-sync, opening the door to AI filmmaking and animated storytelling.
  • Game world creation — build immersive game environments from a single sentence; the model's grasp of space and physics is a notable leap.
  • High prompt accuracy — grounded in real-world physics and backed by strong prompt comprehension for consistent, context-aware results.
  • Cinematic quality — output with smooth motion and realistic effects.

The model was trained by researchers at Google DeepMind and is aimed at creators, developers, and anyone pushing on the limits of AI-generated content.

Prompting Tips

Drawing on Google's own guidance, these strategies sharpen results:

  • Shot composition: "close-up," "two shot," "over-the-shoulder"
  • Lens and focus: "macro lens," "shallow focus," "wide-angle lens"
  • Genre and style: "sci-fi," "romantic comedy," "action movie"
  • Camera motion: "zoom shot," "dolly shot," "tracking shot," "pan shot"

Example prompt:

Close-up shot (composition) of melting icicles (subject) on a frozen rock wall (context) with cool blue tones (ambiance), zoomed in (camera motion) maintaining close-up detail of water drips (action).