Veo Video Generation

Overview

This skill provides seamless integration with Google's state-of-the-art Veo 3.1 video generation model via Vertex AI.

The accompanying script automatically uploads local source images/videos to Google Cloud Storage (GCS), orchestrates the long-running generation task on Google's servers, and downloads the resulting MP4 video back to the local filesystem. It fully supports Veo's advanced features, including creating seamless transitions between two images and extending 8-second clips up to 30 seconds.

When to Use

• Generating high-quality B-roll or cinematic footage from a text prompt.
• Animating a static image (Image-to-Video).
• Creating smooth transitions between two distinct images (First-and-Last Frame interpolation).
• Extending existing Veo-generated videos beyond the initial 8 seconds.
• You need background execution because Veo video generation takes 2-5 minutes per clip.

Don't use for:

• Real-time or instant video generation.
• Scenarios where you lack a configured Google Cloud Project with billing enabled and a GCS bucket.

Environment Configuration

Before running the generation script, ensure the following environment variables are set:

# Google Application Default Credentials
export GOOGLE_APPLICATION_CREDENTIALS="/path/to/your/service_account_key.json"

# Google Cloud Project ID (Optional if inferable from credentials)
export GOOGLE_CLOUD_PROJECT="your-gcp-project-id"

# GCS Bucket for staging inputs and storing outputs
# Ensure the service account has Storage Object Admin rights to this bucket
export VEO_GCS_BUCKET="your-veo-renders-bucket"

# Optional: Override default location (default: us-central1)
export GOOGLE_CLOUD_LOCATION="us-central1"

Modes of Operation

The generate.py script supports four core modes via the --mode flag.

1. Text-to-Video

Generate a video purely from a text prompt (4, 6, or 8 seconds).

python3 ~/.hermes/skills/veo-video/scripts/generate.py \
    --prompt "A cinematic wide shot of a futuristic factory with robotic arms assembling a car, photorealistic, 4k" \
    --mode text \
    --aspect-ratio 16:9 \
    --duration 8 \
    --output /tmp/factory_text.mp4

2. Image-to-Video

Bring a single starting image to life.

python3 ~/.hermes/skills/veo-video/scripts/generate.py \
    --prompt "Smoke starts pouring out of the chimney and the camera slowly pans forward" \
    --mode image \
    --image /tmp/factory_start.jpg \
    --aspect-ratio 16:9 \
    --output /tmp/factory_image.mp4

3. First and Last Frame

Interpolate smoothly from one image to another. Excellent for narrative transitions.

python3 ~/.hermes/skills/veo-video/scripts/generate.py \
    --prompt "The sun sets rapidly and the factory lights turn on" \
    --mode first-last \
    --image /tmp/factory_day.jpg \
    --last-frame /tmp/factory_night.jpg \
    --output /tmp/factory_transition.mp4

4. Extend Video

Extend an existing video by an additional 8 seconds. You can iteratively apply this to reach up to 30 seconds of total runtime.

python3 ~/.hermes/skills/veo-video/scripts/generate.py \
    --prompt "The camera moves closer to the robotic arm as it finishes assembling the part" \
    --mode extend \
    --video /tmp/factory_text.mp4 \
    --output /tmp/factory_extended_16s.mp4

Common Pitfalls

1. Missing IAM Permissions: The script throws an HTTP 403 error.
   • Fix: Ensure the service account has Vertex AI User and Storage Object Admin roles.
2. Bucket Not Found / Access Denied: The upload phase fails.
   • Fix: Verify that VEO_GCS_BUCKET is set correctly and the bucket exists in the same project/location.
3. Timeout / Terminal Blocking: The agent hangs while waiting for the video.
   • Fix: Always run this script using terminal(background=true) since generation takes several minutes.
4. Unsupported Aspect Ratio/Resolution: Image-to-Video yields poor results or fails.
   • Fix: Provide source images perfectly cropped to 1920x1080 (16:9) or 1080x1920 (9:16).

Verification Checklist

• GOOGLE_APPLICATION_CREDENTIALS and VEO_GCS_BUCKET are set in the environment.
• Command is executed in a background terminal to avoid blocking the conversation loop.
• For extend mode, the input video was originally generated by Veo (required by Vertex AI limits).