Is Google I/O’s Biggest Winner Not Google? Inside Gemini Omni Flash

At the May 20 Google I/O event, the company introduced Gemini Omni Flash, a native multimodal large model that accepts video, image, audio, text, or sketch inputs and can generate or edit videos using plain‑language commands while preserving character consistency, physical realism, and contextual memory.

The model embeds an invisible SynthID digital watermark in every generated video, which can be verified through the Gemini app, Chrome, or Google Search. Gemini Omni Flash is currently available to paid users in the Gemini app and Google Flow, will be free in YouTube Shorts and YouTube Create this week, and will be accessible to developers via the Gemini API in the coming weeks.

Early adopters have put the model through rigorous tests. Ethan Mollick, an associate professor at Wharton, used a deliberately complex prompt about a sea otter, Spirit Airlines, and Shakespeare fighting a pizza robot; he noted that the generated video featured smooth shot transitions and high instruction compliance, commenting that “a truly smart model can handle video directly, vastly expanding creative space.”

Other creators reported mixed results. WolfRiccardo produced a fake‑news clip with convincing lighting and realistic expressions, but the final second showed a disappearing phone. Justine Moore tested the model’s world‑knowledge integration by uploading a location photo and asking for its historical background, receiving a detailed answer without extra prompting. Jerrod Lew edited a running clip into varied environments and outfits, preserving motion continuity while changing style. LexnLin demonstrated rapid multi‑angle switching within a 10‑second video, moving fluidly between close‑ups, low angles, and aerial shots. Aimikoda generated a beat‑synchronized portrait showing sixteen precise facial expressions, each hard‑cut on the beat.

Limitations also emerged: the model struggled to faithfully reproduce reference images, exhibited facial consistency glitches (e.g., sudden face swaps), and its detail degraded in wide‑shot frames where facial features became blurry. Users also complained about a strict usage cap—only five videos can be generated before the quota is exhausted, rendering both Pro and Flash modes unusable.

When directly compared with the popular Seedance 2.0, multiple creators used identical prompts. The consensus was that Seedance 2.0 delivered superior facial consistency, smoother and more powerful action sequences, and a more authentic anime‑style rhythm (12 fps “on‑2s” cadence). Gemini Omni Flash’s 10‑second videos leaned toward a 3D aesthetic but suffered from sluggish martial‑arts movements, blurry water‑splashes, and an overall “AI‑generated” feel.

In summary, Gemini Omni Flash shows promise for educational and explanatory video generation and for straightforward editing tasks, but it lags behind competitors in dynamic scene rendering, long‑shot detail, and consistency. Its performance is modestly better than Veo 3.1 yet falls short of external expectations.