Google’s Gemini Omni Flash delivers mixed results in hands-on testing
While the new Flow platform model generates convincing deepfakes, users face steep credit consumption and unpredictable outputs that range from realistic to bizarre.
Google has officially released Gemini Omni Flash, a new generative AI video model accessible through its Flow platform. Positioned as an evolution of its predecessor, Veo, the model claims enhanced character consistency and a deeper understanding of real-world physics. It allows users to generate video content from text, images, and existing video clips, marking a shift toward an "anything-to-anything" generative approach.
Independent testing of the model reveals a performance profile that is both impressive and inconsistent. While some outputs demonstrate a level of realism that surpasses previous iterations, others exhibit significant technical failures. Specific glitches observed include objects changing form mid-clip, such as a jar of honey transforming into a water bottle, and characters suddenly switching orientation during action sequences.
The platform’s editing capabilities, which allow for text-based prompts to modify existing videos, function more effectively than in earlier versions. However, the results can be erratic. Attempts to refine character features, such as removing unintended antlers from a subject, often resulted in the error appearing in other parts of the clip. This suggests that while the model accepts input, its execution remains unpredictable for complex edits.
A more concerning aspect of the testing involves the model’s deepfake capabilities. When used to generate video of the tester’s face in various scenarios, the output was described as highly convincing to untrained observers. Although minor anomalies were present, such as unnatural audio cues and background inconsistencies, the visual fidelity was sufficient to deceive individuals familiar with the subject.
The service operates on a credit-based pricing structure that may deter casual users. Generation costs range from 15 to 40 credits depending on clip length and input complexity, with edits costing 40 credits per round. Under the $20-per-month AI Pro plan, which includes 1,000 credits, usage can deplete the allowance rapidly. Testing showed that approximately 20 clips with edits reduced the monthly credit balance to 145, highlighting the potential for unexpected costs.
The ease of creating realistic fake content has raised concerns regarding the ethical implications of the technology. The tester noted a sense of unease regarding the "uncanny valley" effect, where the barrier between reality and AI generation continues to erode. While the model is not yet capable of producing cinematic masterpieces with trivial effort, its ability to generate plausible video from simple inputs represents a significant step forward in generative AI.
As Google continues to refine its Omni family of models, the focus remains on balancing creative utility with technical reliability. The current iteration of Gemini Omni Flash offers powerful tools for video creation but requires users to navigate a landscape of mixed results and substantial resource consumption.
