Google’s Omni Model Promises ‘Anything-to-Anything’ AI, but the Reality is Still Uncanny
Table of Contents
A New Paradigm for Generative Inputs
At Google I/O 2026, the company unveiled Omni, a family of generative models designed with a conceptually ambitious goal: “anything-to-anything” processing. In theory, Omni is built to take any form of input—be it text, a static image, or a video clip—and transform it into any other medium. While the full scope of this versatility is still being rolled out, the first tangible application has arrived in the form of Omni Flash, now integrated into Google’s AI video editing platform, Flow.
For those familiar with Veo, Google’s previous foray into high-end video synthesis, Omni is positioned as a significant architectural leap. The primary improvement lies in how the model handles reference material. Rather than relying solely on text prompts, users can now upload an existing video to serve as a foundational blueprint, which Google claims results in better character consistency and a deeper integration of real-world physics and knowledge.
The Consistency Gap
Testing Omni’s claims of “consistency” often reveals the persistent gap between marketing and execution. In one exercise involving a recurring character—a stuffed deer named Buddy—the model showed marked improvement over Veo in following complex prompts. However, the “AI jump scares” remain. During a simulated skydiving sequence, the character abruptly shifted orientation mid-air, a glitch that betrays the model’s lack of true spatial awareness.
The model’s struggle with object permanence is even more apparent in narrative clips. In a prompt requesting a montage of the character packing for a cruise, Omni successfully introduced a jar of honey as a plot point. Yet, as the scene progressed, the honey jar mutated into a clear squirt bottle and then back again. While the model managed a coherent narrative beat—having the character mistakenly use honey as sunscreen—the visual instability of the objects themselves remains a hurdle for professional-grade production.
The Cost of Iteration
For users on the $20-per-month AI Pro plan, these inconsistencies create a financial friction point. Generating video in Flow is not free; it consumes a credit system where clips cost between 15 and 40 credits depending on complexity. Edits are particularly expensive, often costing another 40 credits per iteration. When the model misinterprets a prompt—such as adding antlers to a character that shouldn’t have them, and then adding them to every other scene when told to remove them—the cost of “prompt engineering” a usable clip rises quickly.
The Deepfake Dilemma
Where Omni becomes truly unsettling is in its ability to augment real-world footage. Using a neutral selfie video as a base, the model can convincingly map a user into entirely new environments. In tests involving eating spaghetti or standing before the Eiffel Tower, the results were strikingly realistic, bordering on the uncanny.
The synthesis is a high-fidelity blend of generated pixels and original source material. While some auditory tells exist—such as a slightly synthetic clink of a fork against a bowl—the visual integration is sophisticated enough to deceive those closest to the subject. In one instance, the model’s rendering of a person eating pasta was convincing enough to fool a spouse of ten years, who noted only that the bowl looked unfamiliar, even as the facial movements and physics of the action appeared genuine.
This level of realism suggests that while Omni may still struggle to maintain the shape of a honey jar over ten seconds, its ability to synthesize human likeness into believable scenarios is advancing at a pace that far exceeds the current regulatory framework for AI-generated content. As these tools move from the experimental phase in Flow to wider consumer availability, the distinction between a “fun edit” and a high-fidelity deepfake continues to blur.
