Google has unleashed a new generative AI model called Omni, and its capabilities are as wild as they sound. Billed as an anything-to-anything model, Omni is designed to take any form of media — a photo, a video, a snippet of text — and transform it into something else. For now, that means generating video from a combination of uploaded clips and text prompts. But the implications stretch far beyond simple video creation.
The Omni Flash model, the first in this family, is now available in Google's AI video generation and editing platform, Flow. Users can still access the previous Veo model, but Omni brings several improvements. Most notably, you can upload an existing video and use it alongside a text prompt to generate new content. Google also claims Omni has better real-world knowledge, leading to more consistent characters and scenes across a generated video.
Testing Omni: Buddy the Deer Returns
To put Omni through its paces, I resurrected a project from a few months ago: generating videos of my son's stuffed deer, Buddy, on a vacation. The original test with Veo 3 had mixed results — some clips were fun, but others were riddled with inconsistencies. Omni promised better consistency, and in some ways it delivered.
For example, I prompted Omni to create a montage of Buddy packing for a tropical cruise. The model had Buddy pack a jar of honey, and later in the clip, he squirted it onto his hoof like sunscreen. The narrative beat was clever, but the honey container kept changing shape and label. That kind of object inconsistency remains a common failure mode for generative video models.
In another test, I had Buddy go skydiving. The first few seconds were smooth — the deer's tiny parachute deployed and fluttered in the wind. Then, without warning, Buddy flipped upside down, the chute disappeared, and the background warped. It was a classic AI jump scare, the kind that reminds you you're not watching real footage.
Despite these glitches, Omni handled text-based edits much better than Veo. When I asked it to emphasize Buddy's facial reactions, it did produce more expressive eyes and ears — but at the cost of adding antlers, which Buddy does not have. I prompted it to remove the antlers, and it obliged for that scene, only to add them back in every other clip. The model's understanding of negative prompts is still shaky.
Deepfaking Myself: A Troubling Milestone
The most unsettling test involved deepfaking myself. Starting from a simple selfie video with a neutral expression, I asked Omni to generate clips of me eating spaghetti, sitting on a plane, and standing in front of the Eiffel Tower biting a baguette. The results were disturbingly good.
My husband — who has seen my face every day for a decade — watched the pasta-eating clip without knowing it was AI-generated. He believed I was simply eating pasta in front of a camera. His only clue was that the bowl looked unfamiliar. The way my AI self twirled the fork, the slight clink of metal against ceramic, the subtle chewing motions — all felt real enough to fool someone who knows me intimately.
Other deepfakes varied in quality. The airplane clip had a ghost woman in the background, seated twice in the same row. The Eiffel Tower scenes sometimes looked cartoonish, but one had such realistic lighting and hair movement that it would likely pass on social media. The AI version of me turned my head and revealed a ponytail that I don't currently have — but only I would notice that detail.
These results represent a significant leap from Veo. The model's ability to preserve my facial identity, match skin tone, and generate realistic background motion far exceeded my expectations. It took less than five minutes to upload the selfie and type the prompts. The barrier to creating convincing fake videos has never been lower.
The Cost of Creativity
Omni isn't free. Generating videos costs credits, which vary based on clip length and starting materials. A basic 5-second clip costs 15 credits; scenes with multiple input files cost up to 40 credits. Editing an existing clip costs another 40 credits. I subscribed to the $20-per-month AI Pro plan, which includes 1,000 credits. After generating about 20 clips and performing a handful of edits, I had 145 credits left.
If you have a specific vision, be prepared for a costly iterative process. Each edit or regeneration eats into your monthly allowance. For casual users, the credit system may feel restrictive. For professionals, the speed and quality might justify the expense — but the expense can quickly escalate if you're not careful.
Google's pricing also raises questions about accessibility. When generative AI video tools are this easy to use, society must grapple with the potential for misinformation. The company has implemented some safeguards: videos are watermarked, and the model refuses to generate certain content. But as the deepfake test showed, the results can be persuasive beyond what current detection systems can reliably catch.
Where Omni Excels and Where It Falls Short
Omni's biggest strength is its ability to incorporate uploaded video into new creations. You can film yourself sitting in your living room and turn that into a scene at the beach with a simple text command. The model maintains your appearance, clothing, and lighting surprisingly well. Backgrounds are replaced with impressive detail — sand, waves, palm trees all look photorealistic in the best clips.
Character consistency is better than Veo but far from perfect. Buddy kept his spotty pattern in most scenes, but his antlers appeared and disappeared. My own face changed subtly between frames — the shape of my nose shifted slightly, and the color of my shirt varied. The model seems to lose track after a few seconds, especially in longer clips.
Real-world knowledge is a double-edged sword. Omni correctly understood that a cruise ship needs water, that honey is sticky, and that the Eiffel Tower is in Paris. But it also generated a bottle of honey that turned into a squirt bottle and then back into a jar. The model knows facts but still struggles with object permanence.
Ethical and Social Implications
The ease with which Omni creates convincing deepfakes is alarming. In the hands of bad actors, such tools could fuel disinformation campaigns, harassment, or fraud. Even as a tech reviewer, I felt a sense of unease watching my AI doppelgänger eat pasta — a video that, with minimal effort, could pass as authentic to most viewers.
Google has published guidelines for responsible use, and the Flow platform includes content moderation filters. But filters can be circumvented, and watermarks can be cropped or obscured. The company has also restricted access to the model's full capabilities, but similar open-source models are already emerging. The cat-and-mouse game between generative AI and detection tools will likely continue for years.
On the positive side, Omni opens up creative possibilities for filmmakers, educators, and hobbyists. The ability to quickly generate B-roll, change backgrounds, or insert animated elements into real footage could reduce production costs and democratize video storytelling. But with that power comes responsibility — and potential regulation.
The Uncanny Valley Is Home
After spending hours with Omni, I'm left with mixed feelings. The technology is undeniably impressive. It improves on Veo in meaningful ways, offering better editing support and more consistent outputs. Yet the glitches remind you that it's still a statistical model, not a true creator. A deer that skydives upside down, a woman who duplicates in the background, a jar that morphs into a squeeze bottle — these are signs of a system that doesn't truly understand the world.
But the deepfakes are a different story. They cross a threshold that makes me pause. When your own spouse can't tell if a video of you is real, something has fundamentally changed. We are no longer speculating about the potential for AI-generated video to deceive — we are living through the early days of that reality.
Google's Omni is a powerful tool, but it's also a harbinger of the challenges ahead. The line between real and generated will continue to blur, and society will need new norms, technologies, and legal frameworks to navigate that uncertainty. For now, the uncanny valley is our home, and we'd better get used to it.
Source: The Verge News