Google’s new anything-to-anything AI model is wild
Google’s new Gemini Omni video model can edit and generate strikingly realistic clips, but the results are still glitchy and expensive.
The Verge tests Google’s new Gemini Omni video tools and finds them both impressive and unsettling. They can make convincing fake videos from simple prompts, but the model still produces odd visual mistakes and burns through credits quickly.
Google made a new tool that can turn ideas and videos into new videos.
It is a bit like having a magic toy camera that can make a deer go on vacation or make a person look like they are eating pasta in Paris, even if they are not really there.
The catch is that the magic is not perfect. Sometimes the videos glitch or look strange, and making them can use up a lot of points, like a game that costs coins every time it is played.
Omni’s promise
The Verge says Google’s Omni is the newest family of generative models in its Gemini lineup, meant to eventually turn many kinds of inputs into many kinds of outputs. For now, the model being tested, Omni Flash, is focused on video generation inside Google’s Flow tool. It can use an uploaded video plus a text prompt, and Google says it should understand the real world better and keep characters more consistent than earlier tools.
Better, but still weird
To test that claim, the writer returned to an AI-generated stuffed deer character and compared Omni with Google’s previous Veo model. Omni does seem stronger in some areas, especially when editing existing clips or following prompt instructions. But the results remain uneven. The article describes jump-scare style errors, such as a character suddenly changing direction while skydiving, and visual inconsistencies like an object changing shape across shots.
Deepfakes feel more real
The most striking part of the test is how convincing Omni can be when the writer deepfakes herself into scenes like eating spaghetti, sitting on an airplane, and standing at the Eiffel Tower. The clips contain telltale flaws, but they were believable enough to fool her husband when he did not know AI was involved. The piece argues that this is not quite the singularity, but it is a clear step deeper into the uncanny valley.
Cost and friction
The article also notes that the tool is not cheap. Video generation and edits consume credits, and the writer’s quota dropped quickly after a batch of clips. That makes experimentation costly, especially for users trying to get a clip close to what they imagined. The broader takeaway is that AI video is becoming easier, more realistic, and more accessible, while still being imperfect and potentially risky.
Key points
- Google’s Omni Flash can generate and edit video from prompts and uploaded clips.
- The Verge found it more capable than older Veo tests, but still glitchy.
- The model produced convincing deepfake-style clips of the writer herself.
- Some scenes had visual inconsistencies, like changing objects and odd character movement.
- Video generation costs credits quickly, making experimentation expensive.
