Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.ai-stats.phaseo.app/llms.txt

Use this file to discover all available pages before exploring further.

Use this recipe when the video should be guided by one or more images instead of relying on prompt text alone.

​
Goal

  • Send a valid input_references array.
  • Use one image as a first frame or general reference.
  • Keep the request portable across providers where possible.

​
1. Start with one reference image

The video API accepts structured input references. 
curl https://api.phaseo.app/v1/videos \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "google/veo-3.1-lite",
    "prompt": "Animate this image into a slow cinematic push-in.",
    "input_references": [
      {
        "type": "image_url",
        "role": "first_frame",
        "image_url": {
          "url": "https://example.com/reference.png"
        }
      }
    ]
  }'

​
2. Use the role deliberately

Common roles include:
  • first_frame
  • last_frame
  • reference
  • source
Keep the first version simple and use one image first. Then add more references only when the model and provider path clearly support the extra structure you need.

​
3. Keep the prompt aligned with the image

Reference images work best when the prompt explains the intended motion or transformation instead of re-describing the whole scene from scratch. Better pattern:
  • describe the motion, pacing, framing, and ending state
Weaker pattern:
  • re-list every visible object already present in the image

​
4. Verify the selected model actually supports this flow

Before shipping:
  • check the model in GET /v1/videos/models
  • confirm the input types and supported parameters match your request
This avoids building against a prompt-only model and discovering the mismatch after paid requests start failing.

​
5. Add provider-specific overrides only if needed

If one provider needs extra video options, add them through provider_params instead of rewriting the whole request around one provider.
Last modified on May 19, 2026