Introducing Gemini Omni

Summary

Gemini Omni is a new multimodal model family that combines Gemini's reasoning with generative creation, starting with high-quality video output. The first model, Gemini Omni Flash, accepts images, audio (voice references initially), video and text as combined inputs to generate or iteratively edit videos via natural-language conversation. Omni aims for more realistic results by incorporating physics-aware reasoning and grounding in real-world knowledge.

Key Points

• Multimodal inputs: accepts text, images, video and (initially) voice audio as references; can blend references into a single cohesive output.
• Initial output modality: video generation and conversational, multi-turn video editing; image and audio outputs planned in the future.
• Conversational editing: instructions build on previous turns while keeping scene continuity (consistent characters, physics, camera and memory across edits).
• Physics-aware and knowledge-grounded: improved intuitive handling of forces, motion and contextual reasoning to create more realistic and meaningful scenes.
• Style, motion and effect transfer: apply motion from one source to another, transfer visual styles, and compose complex timing/sync to audio.
• Avatars and voice: supports creating videos with a user's digital avatar and voice via Avatars; additional audio editing capabilities are being tested.
• Content transparency: all Omni-generated videos include an imperceptible SynthID watermark; verification available via the Gemini app, Gemini in Chrome and Google Search.

Rollout & developer access

• Gemini Omni Flash is rolling out today to Google AI Plus, Pro and Ultra subscribers via the Gemini app and Google Flow.
• Available at no cost on YouTube Shorts and YouTube Create App starting this week.
• API access for developers and enterprise customers will roll out in the coming weeks.

Practical considerations for engineers

• Design UIs/UX to support multi-turn conversational edits and stateful scene context so edits remain consistent across turns.
• Plan for verification/integrity checks using SynthID when processing or displaying user-generated content.
• Expect expanded input/output modality support (image/audio outputs and broader audio inputs) over time; handle evolving API capabilities and backward compatibility.