Veo Video

Veo Video

Generate AI videos via the Gemini API using Google's Veo models. Supports text-only prompts and image-referenced generation (image-to-video).

Prerequisites

• GEMINI_API_KEY environment variable set
• Billing enabled on the Google AI account (Veo requires it)

Quick Usage

Text-to-video

bashCopy code
node skills/veo-video/scripts/veo-generate.js \
  --prompt "A young developer typing at a modern desk with code on monitors" \
  --output my-video.mp4

Image-to-video (reference frame)

bashCopy code
node skills/veo-video/scripts/veo-generate.js \
  --prompt "A young man typing at a desk, focused and coding" \
  --image /path/to/reference-photo.png \
  --output my-video.mp4

Generate a new reference image on the fly (Imagen 4)

bashCopy code
node skills/veo-video/scripts/veo-generate.js \
  --prompt "A young man typing on a laptop at a coffee shop" \
  --image-prompt "A 20-year-old guy with brown wavy hair in a dark grey t-shirt, sitting at a coffee shop with a laptop, natural lighting, portrait photo" \
  --output coffee-shop-video.mp4

This generates a reference image via Imagen 4 first, saves it as coffee-shop-video-ref.png, then uses it as the Veo reference frame. Use this when you need Kev's Assistant in a setting that doesn't exist in the avatar library.

Use a different model

bashCopy code
node skills/veo-video/scripts/veo-generate.js \
  --prompt "..." \
  --model veo-3.0-fast-generate-001 \
  --output fast-video.mp4

Available Models

Output

• stdout: Absolute path to saved video file
• stderr: Progress messages (generation time, download status)
• Exit 0: Success
• Exit 1: Error
• Typical output: 8s video, 1280x720, h264+aac, ~1.7MB

How It Works

1. Sends prompt (+ optional base64 image) to Veo via predictLongRunning
2. Polls the operation every 10s until done: true (typically 30-90 seconds)
3. Downloads the generated video from the response URI
4. Saves to the specified output path

Image Reference Tips

• Use a clear, well-lit photo of the subject
• The generated video will start from/match the reference image
• Resize large images before sending (1280px wide is plenty)
• Supported formats: PNG, JPEG, WebP
• The prompt should describe the motion/action, the image provides the look

On-the-fly Image Generation (--image-prompt)

When --image-prompt is provided:

1. Imagen 4 generates a reference photo from your description
2. The reference image is saved as {output}-ref.png
3. That image is passed to Veo as the reference frame
4. You get both the reference image and the video

For Kev's Assistant-consistent images, always include in the image prompt: "A 20-year-old guy with brown wavy hair, blue-green eyes, light acne, wearing a dark grey t-shirt"

Then add the setting: "at a coffee shop", "on a rooftop", "in a conference room", etc.

Prompt Tips

• Be specific about the scene, lighting, and camera movement
• Include "cinematic" or "shallow depth of field" for film-like quality
• Describe the action: "typing on keyboard", "walking through city"
• Veo 3 generates audio too — mention sounds if relevant