Google has unveiled a new generative AI model called Omni, which it bills as an anything-to-anything system capable of turning any kind of input—text, images, video—into any desired output. For now, the first available version, Omni Flash, focuses on video generation and editing. It is accessible through Google's Flow platform, replacing the previous Veo model while offering significant improvements in realism and multi-modal input.
The promise of Omni is that users can upload a video clip and combine it with a text prompt to create entirely new scenes. Google claims the model incorporates more real-world knowledge than its predecessor, allowing for better character consistency and more plausible motion. To test these claims, one reviewer brought back a stuffed deer named Buddy, originally used in an earlier experiment with Veo, and subjected him to a series of AI-generated adventures.
The results were mixed. Some clips succeeded remarkably well: Buddy appeared to skydive, pack a suitcase, and board a cruise ship with a degree of coherence that would have been impossible with prior models. Yet the same clips also featured classic AI glitches. During a skydiving scene, Buddy suddenly switched orientation mid-fall. In a montage where he packed honey for a tropical vacation, the honey bottle changed shape and color from one frame to the next—from a jar to a clear squirt bottle and then back to a squeeze bottle. The model even gave Buddy antlers despite the real stuffed animal being a baby deer without them. Removing the antlers in one scene caused the model to add them in all others.
Beyond the stuffed animal tests, Omni proved disturbingly good at creating deepfakes of real people. Using a simple selfie video as a starting point, the model generated clips of the reviewer eating spaghetti, sitting in an airplane seat, and standing in front of the Eiffel Tower holding a baguette. When shown to the reviewer's husband, the pasta-eating scene was convincing enough that he did not suspect anything was generated by AI—only that the bowl looked unfamiliar. The deepfakes contained subtle tells: a fork hitting a bowl sounded too manufactured, and a background woman appeared twice in the airplane scene. But overall, the realism was high enough to fool someone who sees the subject every day.
Omni is not free. It uses a credit system: generating videos costs between 15 and 40 credits depending on length and input, while editing a video costs 40 credits. The basic AI Pro plan, priced at $20 per month, includes 1,000 credits. After generating about 20 clips and performing a few edits, the reviewer was down to 145 credits remaining. Users with specific creative visions may need to spend significant credits iterating with the model to achieve a satisfying result.
Google describes Omni as a family of generative models that will eventually support any input-to-any-output conversion. The current version focuses on video, but the company envisions future capabilities including audio, 3D models, and more. The technology builds on Google's long history in AI research, including the development of transformer architectures that underpin modern large language models and multimodal systems. Omni represents a direct attempt to compete with other video generation tools from companies like OpenAI (Sora) and Meta, while integrating tightly with Google's ecosystem of cloud services and consumer products.
The implications of such realistic video generation are significant. Deepfakes have already raised concerns about misinformation, identity theft, and the erosion of trust in visual media. While Omni includes some safeguards—such as watermarking and content moderation policies—the ease with which the reviewer created convincing impersonations suggests that these protections may not be sufficient. The model could be used for legitimate entertainment, education, and creative projects, but also for malicious purposes like impersonation, fraud, or propaganda.
Critics have pointed out that even the best AI-generated videos remain in the uncanny valley, with subtle inconsistencies in lighting, texture, and physics that give them an unnatural feel. Yet the rate of improvement is accelerating. Veo 3, released only months earlier, produced noticeably poorer results. Omni's ability to accept video input as a starting point marks a notable advance, allowing users to add AI-generated elements to real footage rather than generating entire scenes from scratch. This hybrid approach can produce more credible results, as seen in the deepfake examples.
Google has positioned Omni as a tool for creators, enabling them to produce high-quality video without traditional equipment or expertise. The company has also stressed its commitment to responsible AI development, including transparency measures and collaboration with academic researchers. However, the same features that make Omni useful for benign projects also lower the barrier for creating misleading content. The model can generate a video of a person doing or saying something they never did with minimal effort and without requiring specialized skills.
In the broader context of AI development, Omni illustrates a trend toward increasingly capable multimodal models that blur the line between real and synthetic media. OpenAI's GPT-4 and Google's own Gemini have demonstrated similar abilities in text and image generation. Video generation is the next frontier, and Omni shows that the gap between professional video production and AI-generated clips is closing quickly. The technology is not yet perfect—it still struggles with object permanence, consistent character appearance, and complex physical interactions—but each iteration brings it closer to seamless integration.
For now, Omni remains a tool with both promise and peril. It can transform a simple selfie into a travel video, or a stuffed deer into a vacation star. But it can also create convincing fakes that are difficult to distinguish from genuine footage. As the model improves, the challenge for society will be to develop norms, regulations, and detection methods that allow its benefits while mitigating its harms. Google's Omni is not the singularity, but it is a powerful step into a future where anything-to-anything AI is no longer a distant dream but a present reality.
Source: The Verge News