Google’s Gemini Omni makes realistic video from anything—photos, audio, or text

What happened

At Google I/O 2026, Google unveiled Gemini Omni, an “anything-to-anything” generative model that can synthesize realistic video from nearly any input—images, text clips, existing video, or audio. According to Google DeepMind, the system is designed for conversational, step-by-step editing inside Google Flow, where each instruction builds on the last to preserve a coherent scene rather than generating disconnected clips. The company describes the workflow as “Nano Banana, but for video,” emphasizing that users can refine aesthetics, actions, and effects through natural language at any point.

The capabilities were put to the test in a hands-on review by The Verge’s Allison Johnson, who used the tool to create convincing footage of her child’s stuffed deer rafting and to place herself in a deepfaked scene at the Eiffel Tower. She reported that the process required “surprisingly little effort and know-how,” underscoring how far consumer-grade video generation has progressed in just a year. DeepMind notes that Omni combines an “intuitive understanding of physics” with Gemini’s broader knowledge base to improve narrative consistency and visual realism.

Why it matters

Omni’s multimodal design collapses the traditional pipeline for personalized video. Users no longer need specialized software, motion-capture rigs, or editing expertise to animate a physical object, alter an existing clip, or synthesize a scene from a voice memo and a few sentences. The model’s conversational interface allows iterative refinement, meaning a user can start with a photo, ask for a setting change, then adjust lighting or action in subsequent prompts—making high-fidelity synthetic media accessible to mainstream consumers.

That accessibility is precisely what worries observers. The Verge’s reviewer noted the experiment made her rethink the boundary between harmless creative play and problematic synthetic media, concluding that the tools are now “surprisingly good” at producing realistic output with minimal friction. As personalized deepfakes become trivial to generate from any reference material, platforms and policymakers face renewed pressure to keep pace with detection, disclosure, and consent standards.

Public reaction

No strong public signal was available from Reddit or broader social discussion at the time of publication.

What to watch

Google has yet to publish detailed safety guardrails, API pricing, or commercial licensing terms for Omni outputs. Key questions include how the company will moderate non-consensual depictions of real people, whether generated videos will carry persistent watermarks or content credentials, and how well the model’s physics and scene-consistency claims hold up in third-party stress tests across longer narratives. The speed at which these policies arrive may determine whether Omni is adopted as a creative utility or flagged as a misuse risk.

Sources

• Google’s new anything-to-anything AI model is wild | The Verge — Allison Johnson’s hands-on review
• Gemini Omni — Google DeepMind — Official capabilities, Google Flow access, and model features