Google has introduced Omni, a new family of generative AI models that promises to transform any type of input—text, images, videos—into any other type of output. The first model available, Omni Flash, is now live in Google’s Flow platform, marking a significant step forward in AI-generated video creation. While the concept of an anything-to-anything model is ambitious, the reality in its current form is a mix of impressive realism and frustrating inconsistencies.

Omni builds on Google’s previous Veo model, which already allowed users to generate videos from text prompts. However, Omni adds the ability to upload a video alongside a text prompt, enabling more controlled and contextual generation. Google claims the model incorporates greater real-world knowledge, leading to better character consistency throughout a video. To test these claims, I conducted a series of experiments using a stuffed deer named Buddy and then moved on to deepfakes of myself.

Testing with Buddy the Deer

My first test involved recreating a scenario from a past Gemini ad: making Buddy appear to go on vacation. Using Omni, I prompted the model to generate a montage of Buddy packing for a cruise. The AI created a sequence where Buddy packs a jar of honey, later mistaking it for sunscreen. The narrative was clever, but the video suffered from inconsistencies. The honey bottle changed form multiple times—from a jar to a clear squirt bottle to a squeeze bottle—and the final frame seemed to randomly regurgitate earlier elements. Despite these glitches, the animation was fluid and the character remained recognizable, a marked improvement over Veo.

Another test involved skydiving. Omni handled the motion reasonably well, but Buddy’s orientation would suddenly flip, revealing the model’s struggle with physical consistency. These “AI jump scares” are common in generative video, but they underscore the gap between human-like understanding and statistical pattern matching.

Deepfaking Myself: A Step Closer to the Uncanny Valley

Perhaps the most startling test was deepfaking myself. I provided a short selfie video with a neutral expression and asked Omni to generate videos of me eating spaghetti, sitting in an airplane seat, and posing at the Eiffel Tower with a baguette. The results were disturbingly good. In the pasta video, the clink of the fork against the bowl sounded synthetic, and a background woman duplicated in the airplane clip, but otherwise the scenes were believable. My husband, who sees me daily, was convinced I was actually eating pasta. Only the unfamiliar bowl gave him a hint that something was AI-generated.

The Eiffel Tower clip was even more convincing. The AI me turned her head, revealing a ponytail that contradicted my real appearance, but this subtle detail might escape casual viewers. The model’s ability to insert me into a foreign environment with minimal artifacts is a leap in realism. It took less than a minute to generate each clip, and the editing process—using text prompts to adjust facial expressions or remove antlers from Buddy—worked better than with Veo, though it often introduced new errors.

The Cost of Creativity

Creating these videos isn’t free. Omni uses a credit system: generating a 10-second clip costs 15 to 40 credits, depending on the input complexity, and editing a video costs 40 credits. The $20 monthly AI Pro plan grants 1,000 credits. After generating 20 clips and a few edits, I was down to 145 credits. For users with specific visions, the back-and-forth with the model to achieve desired results can quickly consume credits. This pricing model makes extensive experimentation expensive, limiting access to casual users or professionals with deep pockets.

Ethical and Societal Implications

The realism of Omni’s deepfakes raises urgent ethical questions. If a tool this accessible can produce convincing videos of real people doing things they never did, the potential for misuse is vast. Misinformation, identity theft, and fraud could become easier. Unlike traditional deepfake tools that require technical expertise, Omni simplifies the process to a few prompts. While Google has safety measures, such as watermarking AI-generated content, these measures may not be sufficient to prevent harm. The author’s husband was fooled by a simple dinner video—imagine the impact on a wider audience.

In the broader context of AI development, Omni is part of a trend toward more intuitive generative models. Google’s I/O 2026 announcements included several such advances, but Omni’s anything-to-anything promise is particularly ambitious. It reflects a move away from task-specific models toward general-purpose creative tools. However, the current limitations—inconsistency, glitches, high costs—suggest that a true singularity, where AI seamlessly handles any media transformation, is still some ways off.

When comparing Omni to competitors like OpenAI’s Sora or other video generation models, the key difference is Google’s integration with Flow and the ability to start from a video rather than just text. This gives Omni an edge in context-awareness. The model’s improved character consistency is notable, but it still fails on physical laws and object permanence. For example, Buddy’s honey bottle morphing distracts from the otherwise charming narrative.

The author’s experience reflects a broader user sentiment: amazement at the technology’s capability tempered by frustration with its inconsistency. The phrase “deep in the uncanny valley” is apt. The videos are almost real, but not quite, and that gap can be more unsettling than obviously fake content. As the model improves, the valley will shrink, but the ethical and societal challenges will only grow.

Ultimately, Omni represents a powerful tool for creators, marketers, and everyday users. It can turn a simple idea into a video in minutes, with a level of quality that was unimaginable just a year ago. But with that power comes responsibility. The cost in credits and the potential for misuse are barriers that Google must address. For now, Omni Flash is a wild ride—impressive, flawed, and thought-provoking. It pushes the boundaries of what AI can do, but reminds us that we’re not quite at the horizon of a post-human creative era. The journey continues, and the next model will likely close the gap further, bringing us closer to the anything-to-anything future Google envisions.

Source: The Verge News