What Gemini Omni Is and How It Fits with Gemini 3.5
Gemini Omni is Google’s multimodal AI model that can take mixed inputs like text, images, audio, and video and turn them into coherent, editable video clips grounded in its real‑world knowledge and scene understanding. Announced at Google I/O 2026 alongside the Gemini 3.5 family, Omni sits next to but distinct from Gemini 3.5 Flash, which focuses on frontier intelligence for agents and coding. Where Gemini 3.5 features aim at long, complex tasks and software workflows, Gemini Omni is built for creative and visual work, starting with video generation and editing. The first release, called Gemini Omni Flash, appears in the Gemini app, Google Flow, and YouTube Shorts, placing Google’s video generation AI directly inside products that millions already use. Together, these models form a portfolio that spans creative media, general reasoning, and task‑oriented automation under one Gemini umbrella.
Multimodal AI Capabilities: From Any Input to Video
Gemini Omni’s headline promise is to “create anything from any input,” which means it can start from a photo, a rough sketch, an existing video, or a text prompt and produce new video. This multimodal AI capability is not limited to simple text‑to‑video prompts; Omni can combine images, audio, video, and text in a single request, then generate clips that align with that mix. For example, a user can upload a still photo with a drawn drone flight path and ask Omni to turn it into drone POV footage that follows that path. Another demo applies the pose and motion from one video to a character from a separate image, along with the style of a reference picture. In each case, the model interprets structure from the visual inputs and uses its world knowledge to output footage that feels physically plausible and narratively consistent.
Conversational Video Editing and Scene Memory
One of the most distinctive Gemini Omni video AI features is conversational editing: you talk to the model as if you were giving notes to an editor. According to Google, “every instruction builds on the last,” which means the model keeps track of prior changes and maintains continuity across shots. You can start with a generated or uploaded video, then refine it step by step: change a sculpture into bubbles, swap a background, or adjust camera motion while keeping characters and props consistent. Omni remembers what was visible in previous scenes and tries to keep physics believable, such as a marble rolling smoothly along a chain‑reaction track. This conversational loop turns video into a living document that can be revised through plain language rather than timelines and keyframes, lowering the skill barrier for complex visual storytelling.
Real‑World Demos: From Stuffed Animals to Deepfakes
Public demos highlight both the charm and the risk of Google’s video generation AI. Former Google product manager Bilawal Sidhu fed Omni a photo marked with a drone path, prompting it: “turn this into realistic footage, using the drawing only as a guide for movement, do not show the drawing in the final video.” The model produced drone‑style POV clips that followed the sketched route. The Verge’s Allison Johnson used a photo of her child’s stuffed animal, Buddy, and sent him on AI‑generated adventures like white‑water rafting and snowboarding. She reports that some clips were much more consistent than Google’s earlier Veo tests, yet still had odd “AI jump scares,” such as Buddy suddenly flipping orientation mid‑skydive. Johnson also notes that one deepfake video of herself was convincing enough to briefly fool her husband, showing both the power and the unsettling potential of such tools.
How Gemini Omni Competes and What Comes Next
Gemini Omni’s video skills position it squarely against other advanced AI models focused on generative video and multimodal AI capabilities. Its edge lies in the tight coupling of understanding and generation: it does not only render footage from prompts, but also tracks characters, physics, and scene history across conversational edits. With Gemini Omni embedded in products like the Gemini app, Google Flow, YouTube Shorts, and YouTube Create, creators can turn sketches, photos, or rough cuts into polished sequences without traditional editing tools. At the same time, Google is marking all Omni‑generated videos with its imperceptible SynthID digital watermark so that content originating from its ecosystem can be identified later. This combination of accessible creativity, agent‑class intelligence from Gemini 3.5, and watermarking suggests a future where AI‑driven video becomes routine, while questions about misuse, authenticity, and “net benefit to society” will remain central.
