Google’s Omni Model Promises ‘Anything-to-Anything’ Generation, But the Reality is More Chaotic
Table of Contents
The Ambition of ‘Anything-to-Anything’
Google is pivoting toward a future where the boundaries between media types are virtually non-existent. At the center of this shift is Omni, a new family of generative models designed to handle ‘anything-to-anything’ transformations. In theory, Omni could take a snippet of text and turn it into a video, or take a photograph and morph it into a fully realized cinematic sequence. For now, however, the rollout is focused on the video front.
The first iteration, Omni Flash, has arrived within Google’s AI video generation and editing platform, Flow. While users can still opt for the previous Veo model, Omni is positioned as a significant leap forward. Google claims the new model possesses a deeper grasp of real-world physics and a vastly improved ability to maintain character consistency across different scenes—two of the most persistent hurdles in generative video.
Testing the Consistency Gap
To put these claims to the test, we pushed Omni to handle a complex narrative involving a consistent character: a child’s stuffed deer. The goal was to see if the model could maintain the toy’s appearance across a series of diverse environments, from skydiving to a tropical cruise.
The results highlight the current state of the ‘AI uncanny valley.’ In some instances, Omni outperformed Veo, producing clips that were remarkably true to the prompt. However, the stability is fragile. During a skydiving sequence, the character suddenly flipped orientation mid-air—a classic ‘AI jump scare’ where the model loses track of spatial logic.
The struggle with object permanence was even more evident in a sequence involving a jar of honey. Throughout a single montage, the honey container morphed unpredictably from a glass jar to a clear plastic squirt bottle and back again. The clip ended in a surrealist blur, as if the model had simply collapsed the remaining visual elements of the scene into a single, incoherent frame.
The Cost of Iteration
Beyond the technical glitches, there is the matter of the economy behind these tools. Omni is not a free experiment; it operates on a credit system. Generating a scene can cost anywhere from 15 to 40 credits depending on the complexity and length. A single round of edits—attempting to fix the very glitches the model introduces—costs another 40 credits.
For users on the $20-per-month AI Pro plan, which provides 1,000 credits, the burn rate is surprisingly high. After generating roughly 20 clips with a few iterative edits, the credit balance plummeted to 145. This creates a frustrating paradox: the model often requires significant back-and-forth to achieve a usable result, but the cost of that iteration quickly exhausts the monthly allowance.
The Deepfake Dilemma
Where Omni becomes truly unsettling is in its ability to integrate AI-generated elements into real-world footage. By uploading a neutral selfie video, we prompted the model to place us in various scenarios: eating spaghetti, sitting on a plane, and visiting the Eiffel Tower.
The results were frighteningly convincing. While there are still ‘tells’—a manufactured sound of a fork hitting a bowl or a recurring background extra in the airplane clip—the visual fidelity is high enough to deceive. In a blind test, a spouse who has spent a decade with the subject was unable to tell which parts of a pasta-eating clip were synthetic, noting only that the bowl looked unfamiliar.
This level of realism moves Omni out of the realm of ‘harmless fun’ and into a more complicated ethical space. When a tool can convince someone who knows your every mannerism that you are in a place you’ve never been, the line between a creative tool and a disinformation engine becomes perilously thin.
