Gemini Omni Flash 1.2 API

Gemini Omni Flash 1.2 is the newest version of Google's Omni video model, the system that folds Gemini's multimodal understanding and Google's generative media work into one model that both makes and edits video. Google calls it Gemini Omni 1.2 Flash and announced general availability on 5 November 2026. It writes clips with native sound from a text prompt, from up to six reference images, from a first frame or a pair of first and last frames, and it can take an existing video and change it.

The headline change is length. Omni 1.1 stopped at 10 seconds; 1.2 goes to 15, and any whole second from 3 to 15 is allowed. That is enough for a full social spot, a complete product turn or a short exchange of dialogue in a single take. The quieter changes may matter more: Google says 1.2 holds faces, garments, packaging text and brand colours steady across camera angles, routes dialogue to the right speaker with better lip sync, and keeps camera moves continuous instead of breaking them with jump cuts.

On Unicorn API the model is available as google-gemini-omni-flash-1-2, one id that covers every mode. Which inputs you send decides what it does.

Where Gemini Omni comes from

Omni is Google DeepMind's answer to a question the Veo line left open: what happens when the video model is also a Gemini model. Google introduced the first Gemini Omni Flash at Google I/O on 19 May 2026, describing it as a model that combines Gemini's intelligence with its generative media models. It reached consumers first, through the Gemini app and YouTube, where it replaced Veo 3.1 as the default video model for paying Gemini subscribers. Developers got API access at the end of June. Veo did not disappear; it remains in Google's Flow filmmaking tool, and the two lines now sit side by side, with Omni positioned as the one built around references, reasoning over mixed inputs and conversational editing.

Gemini Omni 1.1 Flash followed on 27 August 2026 as the generally available successor to the preview. It added the controls developers had asked for: 1080p and 4K output by upscaling, a fast 360p draft resolution, first-and-last-frame interpolation, scene extension and video input for edits. Its clips ran from 3 to 10 seconds depending on the task.

Version 1.2 keeps that toolset and works on length and reliability. Google is retiring the old id: requests to Gemini Omni 1.1 Flash in its own API are routed to 1.2 from December.

Like every Gemini model, Omni is closed. Google has published no weights, no parameter count and no architecture paper, and use is governed by the Gemini API terms. The model is reachable through the Gemini API, Google AI Studio and Google Cloud, through Google products such as Vids, Flow, the Gemini app and Google Ads, and through partners such as Unicorn.

What changed from Omni 1.1 Flash

Fifteen-second clips

The 10-second ceiling of 1.1 is now 15 seconds, a 50 percent increase, and the length is chosen in one-second steps from 3 upward. For text prompts, reference images and a single start frame, the full 3 to 15 second range applies. When you pin both a first and a last frame, the range stays at 3 to 10 seconds.

Reference details that hold from start to finish

Google's second claim is reference adherence. Give the model a face, an outfit, a bottle with a label or a brand palette, and 1.2 is meant to keep those details intact from the first frame to the last, even as the camera moves around the subject. Google pitches this directly at virtual try-on and branded advertising, where a shirt that changes its pattern or a label that turns to gibberish ruins the shot. Testers also report that 1.2 matches the layout and text of a reference image more closely than 1.1 did.

Steadier characters, objects and scenes

Consistency is the one improvement Google put a number on. In its internal comparison, raters preferred 1.2 over 1.1 for visual consistency 62 percent of the time. No independent benchmark or public leaderboard score accompanies the release, so that figure is Google's own.

Dialogue between several speakers

Omni generates sound in the same pass as the picture, and 1.2 improves the part that was weakest in early reviews: people talking. Reviewers of the first Omni Flash found that single-speaker lip sync drifted after six or seven seconds and that multi-person scenes often gave a line to the wrong character. Google says 1.2 handles multi-speaker dialogue better, with distinct voices routed to each character and tighter lip sync.

Camera moves without jump cuts

Finally, Google describes more controlled, cinematic camera work: flythroughs, orbits and slow zoom-outs that hold as one unbroken move, with fewer of the sudden cuts that earlier versions inserted mid-shot. Early demonstrations lean on exactly this, with drone paths through buildings and first-person routes across changing terrain.

How Omni works

Google describes the architecture only at the highest level. The model card says Omni combines Gemini's intelligence with Google's generative media models and accepts text, images, audio and video as input, producing high-resolution video with audio. The practical consequence is that the model reads all inputs together: a prompt, several pictures and a clip are interpreted as one instruction rather than handled by separate tools. That is also why Omni is good at explainers that need world knowledge, such as a diagram of a process that should be correct, not just pretty.

Two technical details are worth knowing. The 1080p and 4K outputs are upscaled from the model's native render, with 720p as the default. And outputs carry SynthID, Google's invisible watermark, so they can be identified as AI-generated. Google has not published a frame rate.

Gemini Omni Flash 1.2 specs on Unicorn

The model has four ways in, selected by the fields you send:

• Text and references. A prompt, optionally with up to six images in image_urls for a subject, scene, style or storyboard. Two reference images are not read as start and end frames; they are references.
• First frame. first_frame_url sets the opening image and the model continues from it, 3 to 15 seconds.
• First and last frame. Add last_frame_url and the model interpolates between the two, 3 to 10 seconds. This mode takes no reference images or video.
• Video edit. One URL in video_list switches to editing. Leave out duration; the result runs as long as the source. Reference images can be added to show what should appear.

With aspect_ratio set to auto, the model follows the reference image; without one, auto means 16:9. A fixed seed makes reruns comparable when you change a single word of the prompt.

Where Omni 1.2 shines

Brand and product work. Six reference images and Google's focus on packaging typography and brand colours make it a strong first choice for ads where the product must look exactly like the product, from every angle.

Virtual try-on and recurring characters. A face and a garment supplied as references should survive a turn, a walk and a change of camera angle, which is what try-on and character-led series need.

Single-take camera moves. Fifteen seconds of uninterrupted flythrough, orbit or reveal suit property tours, architecture, travel and establishing shots.

Short dialogue scenes. Two or three characters with their own voices, ambience and effects come out in one pass, ready for a social cut.

Editing footage you already have. Background swaps with matching light, removed objects, a new time of day or a different product in the hand. Reviewers of the first Omni singled out this kind of VFX cleanup as its standout skill, and 1.2 keeps it.

Explainers. Gemini's world knowledge helps with scientific and technical scenes that need to be plausible in their details.

Limits to know before you build on it

Fifteen seconds is the ceiling. Anything longer means generating several clips and stitching them. Some competing models produce longer single clips.

Physics and fast action. Google itself lists fast sports action, rapid script changes and complex physics as areas still under development. Earlier reviews found that violent or fast person-to-person contact was damped and that elaborate physics prompts landed about half the time.

Upscaled high resolutions. 1080p and 4K are upscales of a 720p-class render. They are sharper deliverables, not more generated detail.

Language. Google fully supports English prompts and dialogue; other languages have not been validated.

Two shapes. Only landscape 16:9 and portrait 9:16 are available; there is no square or cinema-wide format.

One request, one result. In Google's own apps, Omni's editing is conversational over several turns. On Unicorn each job stands alone: to refine a clip, send the previous result back in video_list with the next instruction.

Guarded features. Google restricts changing what real people say in a video, applies content filters to recognisable people and minors, and watermarks every output with SynthID.

Prompting tips for Gemini Omni Flash 1.2

Google's guidance is to describe the scene, the action, the camera and the sound, in that order. Omni reads prompts as a Gemini model does, so plain, complete sentences work better than tag lists. Name the camera move and say explicitly when it should be one continuous shot. Put dialogue in quotation marks and attach each line to a clearly described character. When you use references, say what each image is for: the face, the jacket, the bottle, the style.

A continuous camera move:

One continuous drone shot with no cuts: the camera glides low over rows of tomato plants inside a tall glass greenhouse, rises past hanging lamps, and pushes out through an open roof panel to reveal green hills at sunrise. Soft morning light, gentle fan hum and birdsong.

A two-person dialogue:

Two friends sit on a park bench in autumn. The woman in the yellow scarf says, "You kept the map?" The man with the grey beard pulls a folded paper from his coat and answers, "Every crease of it." Static medium shot, warm afternoon light, leaves rustling.

A product shot with references (send the product photo and the model photo in image_urls):

The model from the second image holds the bottle from the first image at eye level and turns it slowly so the label faces the camera. Slow orbit around her, studio light on a pale blue background. Keep the label text and colours exactly as in the reference. Soft electronic music.

An edit (send the clip in video_list):

Change the time of day to dusk: warm orange sky, street lamps switching on, longer shadows. Keep the people, their movements and the camera exactly the same.

Draft at 360p or 720p with a fixed seed while you tune the wording, then render the keeper at 1080p or 4K.

Gemini Omni Flash 1.2 against other video models

Kling 4.0 Reference To Video takes more references and runs longer clips. Choose it for long takes built from a large reference set; choose Omni 1.2 when you also need to edit existing footage or want one model for every mode.

Kling 3.0 Omni shares the "omni" idea of one model for generation and editing. Run the same references through both when consistency is the brief.

Seedance 2.5 produces longer single clips and is the comparison to make for narrative pieces that would otherwise need stitching.

Wan 3.0 and Vidu Q4 Reference To Video are the other reference-driven models worth a side-by-side test on the same product or character.

Getting started

Open Gemini Omni Flash 1.2 in the studio, write a prompt, add reference images, a start frame or a clip to edit, and choose length, resolution and shape. For code, create a key on the API keys page and submit a job to /v1/jobs with model set to google-gemini-omni-flash-1-2 and a prompt, plus whichever inputs your mode needs. The API tab on this page has complete request examples, and the docs explain how jobs and results work.

Verdict

Gemini Omni Flash 1.2 is a refinement with real reach. Fifteen-second clips remove the most common reason to stitch, and the work on references, dialogue and continuous camera moves targets exactly what went wrong in earlier Omni versions. It is not the model for long-form scenes or hard physics, and Google's own consistency figure still awaits independent confirmation. But for brand content, try-on, character-led shorts and editing footage you already own, it is one of the most complete single models in the catalog: one id, six references, frame control, native sound and editing, from a 360p draft to a 4K delivery.