The much-hyped Google I/O event did not disappoint. After a bit of preamble, the real highlight: Gemini Omni, a new family of native multimodal, "any-to-any" AI models, which means that they process you name it - text, image, audio, and video simultaneously. In other words, Nano Banana, but for video. (This is how Google itself talked about it.)
The initial model is Omni Flash. It introduces conversational video editing for creating cohesive media from individual assets. So, you can edit and remix videos through conversations and not by using editing software.
What Is Gemini Omni?
Let's understand what Omni is about in more detail.
Gemini Omni is Google DeepMind's first natively multimodal generative media model. The first variant in the family is Gemini Omni Flash, which is now live
- inside the Gemini app,
- Google Flow, and
- YouTube Shorts
Omni accepts any combination of text, images, audio, and video as input and produces a video. Key here: There's no relay happening across different systems; this is all one model.
Until now, Google ran a split stack: Veo for video, Imagen for images, separate systems for audio. Omni collapses that into one model (hence the name!) that can reason across modalities. In practice, that translates to more coherent edits and fewer pipeline artifacts.
Before I get carried away, let's frame this as key features:
Key Features of Gemini Omni
Here's what stands out in terms of features:
Conversational video editing
The headline capability is conversational video editing. You give Omni a clip — one you generated or one you shot on your phone — and you change it by talking to it. "Make the lights dim." "Change the camera angle to over her shoulder." "Turn the violin invisible." Each instruction persists across turns. The scene evolves; it doesn't reset.
Real-world physics and world knowledge
Google made a point of this at launch. Omni has an intuitive grasp of gravity, kinetic energy, and fluid dynamics.
Sketches and drawings to video
You can turn doodles into realistic footage, using the sketch only as a movement guide rather than as a final visual reference. This will be useful pre-production. Kids will love this, for their drawings, also.
SynthID watermarking and C2PA content credentials
Every Omni output ships with two layers of provenance.
- SynthID is an invisible watermark embedded directly into the pixels at generation time. It's imperceptible to viewers. And it's designed to survive cropping, filters, and re-encoding.
- C2PA content credentials sit alongside it as a signed cryptographic manifest attached to the file. Strip the metadata, and the pixel-level signal still holds.
Worth flagging: SynthID is Google-specific. Only Google's own models embed it, so "not watermarked" doesn't mean "human-made" — it just means "not from a Google model." Given how easily Omni can remix real footage, durable provenance on every output starts to look less like a nice-to-have and more like table stakes.
Testing Gemini Omni
Let's give it a test:
Testing physics and world knowledge
Here was my prompt:
A medieval trebuchet launching a fired clay pot at a stone castle wall, shot in slow motion. The counterweight falls, the sling whips around, the pot arcs through the air and shatters against the stone, shards and embers scattering across the courtyard. Continuous handheld camera move, golden hour light, period-accurate construction and dress. Realistic sound design — wood creaking under tension, rope strain, the whoosh of the sling, the sharp crack of impact. No music.
The note told me it would take a few minutes, but it really only took about ten seconds.
