A seasoned tech reviewer with over ten years of experience specializing in consumer technology, particularly in mobile photography and telecommunications, has shared her insights. Previously affiliated with DPReview, she offers a unique perspective on the evolving landscape of generative AI and its implications for creative media.
Last year, in a playful endeavor, I created a deepfake involving my child's stuffed toy, transforming his plush deer into a traveler exploring various adventures. This was inspired by a Google advertisement featuring the Gemini model, and although I never shared these videos with my four-year-old, the experience prompted me to reflect deeply on the nuances between lighthearted experiments with generative AI and more serious applications. There's a grey area worth exploring, especially given how remarkably accessible these technologies have become, particularly with the recent launch of Gemini’s Omni model.
Omni represents a progressive suite of generative models with the ambitious aim of transforming various input types—photos, videos, or text—into others. While it aims to eventually create broader content, it currently focuses on video generation. The first model introduced under this umbrella, Omni Flash, is now part of Google’s AI-driven video generation and editing service, Flow. Users can still opt for the earlier Veo model, but Omni brings improvements that are hard to ignore.
With Omni, you can initiate a creative process by uploading a video alongside a text prompt. According to Google, Omni is designed to leverage more real-world context when generating videos, resulting in improved character consistency. To test this claim, I resurrected AI Buddy for another round of adventure-filled escapades.
My findings were somewhat perplexing; the outcomes were a mixed bag. Several clips exhibited better consistency and alignment with my prompts compared to my previous experiences with Veo five months prior. Yet, some of Omni’s best offerings still featured notable glitches, such as Buddy’s unexpected change in position during a skydiving scene.
For one video, I granted Omni some creative latitude by requesting a montage of Buddy preparing for a vacation. I specified a playful and cute mood, suggesting that Buddy humorously packed an item that would play a role later. The AI had him packing a jar of honey, which he later mistakenly squirted onto his hoof instead of applying sunscreen. While that bit was amusing, the inconsistency was jarring, transitioning from a jar to a bold water squirt bottle, and back again.
The tool allows you to suggest text-based edits to your clips, and I have to commend Google—this aspect functions better with Omni than it did with Veo 3. However, the results can still be hit-or-miss. When I asked it to enhance Buddy's facial expressions in his clips, the outcome became awkward. Additionally, there were moments where Buddy inexplicably sported antlers, which was a glaring inaccuracy given that he is just a baby deer. After asking it to correct the antlers in one scene, it proceeded to add them to all the others.
The service isn’t without cost; generating videos requires credits, ranging from 15 to 40, depending on scene length and complexity, with each round of edits costing 40 credits. My monthly subscription to the AI Pro plan, costing $20 and providing 1,000 credits, has seen me whittle down to just 145 credits after generating around 20 clips and making several edits. If you have specific visions for your videos, be prepared for potentially costly iterations to achieve your desired result.
Trying out Omni's features on myself, I ventured into deepfaking my own image. Starting from a neutral selfie video, I prompted the model to create clips of me enjoying spaghetti, sitting in an airplane, and feasting on a baguette at the Eiffel Tower. The results were surprising, though not without their flaws.
The deepfake videos exhibited clear AI artifacts. For example, the sound of the fork against the pasta bowl felt artificially generated, and a woman appeared multiple times in the airplane scene. Despite these quirks and the overall eeriness, the final results were compellingly realistic.
When I shared the pasta clip with my husband—without revealing which parts were AI-generated—he accepted it as real, noting that the bowl seemed unfamiliar was his sole hint that something was off. The transformation in my other deepfakes varied; some looked cartoonish, while one clip proved sufficiently convincing to require multiple views to identify as AI-generated. Although subtle details, like my hair being tied back, gave it away for me, they might easily evade casual viewers.
I have mixed feelings about the progress in this domain. My previous tests with Veo 3 left me astonished by its realism, and I've faced a continuous series of revelations as the technology advances. With Omni, while I remain impressed, the novelty of such capabilities has worn off slightly.
Producing an AI-generated cinematic masterpiece is still not as effortless as suggested. Yet, Omni indeed refines some elements present in Veo. For Google account holders, it’s now simpler to manipulate a home video to mimic a flight to an exotic locale with minimal effort. While we haven't reached the pinnacle of technological singularity, we are certainly delving deeper into the uncanny valley.
All visual content in this article was generated by Google’s Gemini model.
