google-gemini-omni-video

What it does

The skill is a backend/API oriented capability that enables generation, referencing, and conversational editing of short videos using Google's Gemini Omni Flash. It explicitly instructs calibrating the backend (Developer API or Enterprise Agent Platform) based on contractual constraints, handling 720p video, 3–10 seconds duration, 24 FPS, and allowed aspect ratios. It includes guidance for planning costs, quotas, approvals, and secure handling of source material, with explicit stepwise considerations before creating or editing videos. It also provides concrete Python example scaffolding for interacting with the API, and it outlines how to manage tokens, interactions, and file handling for outputs and state persistence.

How it works

• Establish scope before acting: ask about backend type (Developer API or Enterprise Agent Platform), data region, source media, edit mode, aspect ratio, duration, audio/dialogue, audience, rights, and spend approval.
• Choose backend deliberately: use Developer API for quick integration and API-key state; use Enterprise Platform for.cloud governance and data controls. Note location and region handling differences and verify exact Omni routes when needed.
• Apply capability contract: Developer API outputs a single 720p video with audio, 3–10 seconds, 24 FPS; supports task types like text_to_video, image_to_video, reference_to_video, and edit; new edits create new videos with stored parent interaction; read video content from steps[].type == "model_output"; synthesized output may include invisible SynthID; safety filters apply.
• Current exclusions and conflicts: certain features are not supported (e.g., audio-reference uploads, certain editing features, minors in image uploads in some regions); resolve conflicts between Cloud vs Developer guidance by treating some capabilities as UNKNOWN until confirmed.
• Plan cost and approval: no free tier; explicit cost per tokens and per second of video; require upfront estimate before creates/edits; obtain explicit approvals including backend and numbers of paid creates.
• Secure handling and prompts: enforce rights and consent, hash sources, manage retention, and follow documented handling for uploads and outputs; use provided code samples and maintain safe workflows.
• Build prompts and testing: prompts should specify subject, action, environment, shot type, timing, and sound; include timecodes for direction; provide explicit media roles and references when needed.
• Safe Developer API workflow: prefer URI delivery for outputs, save initial interaction ID/URI, poll status, download securely; handle idempotency hints and retries with caution; use a sample Python workflow illustrating messaging and guardrails.

When to use it

Use when a task specifically requires Gemini Omni video, multimodal reference roles, time-directed prompting, audio-aware generation, Interactions API state, or a precise comparison against other routes like Veo or Nano Banana. Consider it for first-party Gemini Omni video tasks where you need explicit backend and regional governance disclosures and controlled prompts.

What it can touch

• Developer API or Enterprise Agent Platform endpoints for interactions: including REST endpoints and OAuth/IAM tokens.
• Local sample Python code illustrating interaction with the Gemini API and generating a 10-second video output.

Caveats

• Gemini Omni Flash is a preview model; Interactions API is GA, with mixed lifecycle promises across pages; monitor release notes.
• Pricing and quotas are complex and depend on backend; some Cloud availability statements are UNKNOWN pending confirmation.
• Some features are explicitly not supported in the current Developer API (e.g., audio-reference uploads, certain editing features, or regional restrictions in certain jurisdictions).
• Output may include invisible SynthID; C2PA and certain post-processing guarantees are not assured.
• Always verify model endpoints and regional constraints before production use. The plan emphasizes explicit approvals and conservative estimates, and to treat certain metrics as estimates, not caps.