# Gemini Omni Gemini Omni is a multimodal video generation and editing model from [[Google DeepMind]], part of the broader [[Gemini]] family. It brings the conversational, iterative editing style of modern image generation to video — Google's own framing is "like [[Nano Banana]], but for video." ## What it does - **Conversational video editing** — edit by natural-language prompt; each edit builds on the last while scene coherence holds - **Multimodal input** — takes images, text, video, and audio as references and combines them into one scene - **World knowledge plus physics** — pairs an intuitive grasp of physics (gravity, kinetic energy, fluid dynamics) with Gemini's knowledge of history, science, and culture - **Object and motion control** — character and object swapping via reference images, motion transfer, sketch-to-video, style reimagining - **Output** — video with synchronised visuals, text rendering, and motion ## Prompting Google's prompt guide frames a good prompt around six elements: shot framing and motion, style, lighting, location, action, and camera technique (e.g. "oner", "dolly zoom", "push in"). Reference images carry consistency across iterative edits. ## Availability Via the Gemini app, Google Flow, and YouTube Shorts. A Google AI subscription is required; features vary by tier and region, and early access has been tightly rate-limited. ## How it's received Reception is mixed. The conversational editing and physics grounding impress, but practitioners report the hard part is still unsolved: deep spatial understanding. Testers describe geometry that morphs when it leaves and re-enters frame, and objects that disappear or merge under physical stress (a falling Jenga tower was a cited failure). Several found ByteDance's Seedance and OpenAI's [[Sora]] stronger on raw sample quality, and access limits drew as much comment as the model itself. ## Inside Genie 4 Gemini Omni is also a building block of [[Genie 4]] (October 2026), DeepMind's real-time world model: Omni grounds how the simulated world looks and moves, while Genie keeps it interactive. The same weakness shows up there, with objects that change when they leave the frame and come back. ## Versions - **Gemini Omni 1.1**: 10-second clips - **[[Gemini Omni 1.2 Flash]]** (5 November 2026): 15-second clips, better reference adherence and consistency, multi-turn editing on full clips, $0.07 per second of 720p video via the API (30% less than 1.1). The first generally available version. Google says fast sports action and complex physics still need work ## References - Model page: https://deepmind.google/models/gemini-omni/ - Prompt guide: https://deepmind.google/models/gemini-omni/prompt-guide/ - Hacker News discussion: https://news.ycombinator.com/item?id=48196609 ## Related - [[Gemini Omni 1.2 Flash]] - [[Gemini]] - [[Google DeepMind]] - [[Gemini 3]] - [[Nano Banana]] - [[Gemini 3.1 Flash Live]] - [[Gemini 3.1 Flash TTS]] - [[Gemini Mobile App]] - [[Sora]] - [[Genie 4]]