The Gemini Omni Family: The Essentials

Gemini Omni is Google's any-input video family on Scenario, and the current generation is Gemini Omni 1.2 Flash: five sibling models covering generation, editing, continuation and upscaling. Create a clip of 3 to 15 seconds from text or an image at up to 4K, restyle an existing clip in plain English including full style and setting reinvention, drop a specific subject into a fresh scene with spoken dialogue, extend any clip, including outputs from other models, with matching motion and native audio, or upscale a finished clip to 1080p or 4K.

Gemini Omni 1.1 Flash stops working on December 3, 2026, and the original Gemini Omni Flash models stopped working on November 6, 2026. Move your workflows, apps and API calls to the matching 1.2 model, as described under Moving to Gemini Omni 1.2 Flash below. The worked examples in this article were made with Gemini Omni 1.1 Flash, which takes the same inputs as 1.2.

Full art-style transfer: griffin to Studio Ghibli watercolor

Prompt: "Restyle this entire clip as a hand-painted Studio Ghibli-style animation. Preserve the original camera motion and timing exactly. Convert the griffin's photoreal feathers and fur into soft painterly watercolor-style linework with visible brushstroke texture, the rocky cliff and overcast sky becoming a hand-painted pastel background with soft cel-shaded lighting. Keep the griffin's exact pose, wing motion, and the overall composition identical, just change the rendering style."

Before: Photoreal griffin. After: Ghibli watercolor griffin

A full art-style transfer with Gemini Omni 1.1 Flash Edit, motion and pose held exactly, only the look changed.

Which Model Should I Use?

Gemini Omni 1.2 Flash is the current generation: five models covering generation, editing, continuation and upscaling, with native audio, including lip-synced dialogue, throughout. It replaces Gemini Omni 1.1 Flash model for model and adds Gemini Omni 1.2 Flash Upscale.

Moving to Gemini Omni 1.2 Flash

Gemini Omni 1.1 Flash is being retired. Gemini Omni 1.1 Flash, Gemini Omni 1.1 Flash Edit, Gemini Omni 1.1 Flash Reference to Video and Gemini Omni 1.1 Flash Extend stop working on December 3, 2026. Until then, the workflows, apps and API integrations that use them keep working; after that date, generations with them fail. Videos you already generated stay in your library.

The original Gemini Omni Flash models have stopped working. Gemini Omni Flash, Gemini Omni Flash Edit and Gemini Omni Flash Reference to Video were retired on November 6, 2026. If a workflow, app or API integration still uses one of them, its generations now fail: switch it to the matching 1.2 model.

The 1.2 models take the same inputs, so in most cases switching the model, or its model ID in an API call, is enough:

• Gemini Omni 1.1 Flash (model_google-omni-1-1-flash) and Gemini Omni Flash (model_google-omni-flash): use Gemini Omni 1.2 Flash (model_google-omni-1-2-flash) for text to video and image to video.

• Gemini Omni 1.1 Flash Edit (model_google-omni-1-1-flash-edit) and Gemini Omni Flash Edit (model_google-omni-flash-edit): use Gemini Omni 1.2 Flash Edit (model_google-omni-1-2-flash-edit). The source video must be 15 seconds or less.

• Gemini Omni 1.1 Flash Reference to Video (model_google-omni-1-1-flash-r2v) and Gemini Omni Flash Reference to Video (model_google-omni-flash-r2v): use Gemini Omni 1.2 Flash Reference to Video (model_google-omni-1-2-flash-r2v).

• Gemini Omni 1.1 Flash Extend (model_google-omni-1-1-flash-extend): use Gemini Omni 1.2 Flash Extend (model_google-omni-1-2-flash-extend). A prompt is now required.

Compared with 1.1, the 1.2 models make clips of up to 15 seconds (10 with a last frame), take up to 10 reference images, and cost less at 360p and 720p. The family also adds Gemini Omni 1.2 Flash Upscale, which takes a 360p, 720p or 1080p video to 720p, 1080p or 4K.

Text and Image to Video: Examples from Gemini Omni 1.1 Flash

The base model turns a prompt, an optional first frame and an optional last frame into a clip with native audio, from 360p through 4K. Scripted dialogue in quotes lands with natural lip sync, which makes it a strong pick for anything with a speaking character: a spokesperson, a game NPC, or a narrative beat with a line of dialogue. The examples below were made with Gemini Omni 1.1 Flash, which is retiring. Gemini Omni 1.2 Flash takes the same inputs, and its clips run up to 15 seconds.

Ship captain, scripted dialogue

A weathered ship's captain in a navy peacoat stands at the bow of a sailing ship, raises a brass telescope to scan the horizon, then turns to the crew and calls out in a booming voice: "Land ho! Prepare to make port!" The crew erupts into muffled cheering as he turns back to the horizon, coat billowing in the wind.

Settings: Text-to-video, 16:9, 10 seconds, 1080p. Watch: Open on Scenario

Comedic slow motion, no dialogue needed

A golden retriever stands on its hind legs at a backyard grill, a stolen sausage hanging from its mouth. It freezes mid-bite as it hears a sound, ears pinning back, tail sinking guiltily, then backs slowly off the table in exaggerated slow motion, sausage still clenched triumphantly in its jaws.

Settings: Text-to-video, 16:9, 10 seconds, 1080p. Watch: Open on Scenario

Edit: Full Style and Setting Reinvention, Examples from Gemini Omni 1.1 Flash Edit

Edit handles season and object swaps, and goes further: complete art-style transfer and full setting replacement, aiming to keep the source clip's pose and motion. In the two transforms tested end to end below, made with Gemini Omni 1.1 Flash Edit, pose and motion held: a photoreal creature repainted as hand-drawn animation, and a character's entire environment swapped for a different one. Results can still shift the motion, so review each edit before you use it. With Gemini Omni 1.2 Flash Edit, the source video must be 15 seconds or less.

Full setting swap: space station to deep-sea dive

Prompt: "Change the entire setting and the astronaut's suit from a space station module to a deep-sea diving environment, without altering their pose, floating motion, or hand movement toward the droplet. Replace the spacesuit with a similarly proportioned vintage brass diving suit and round porthole helmet, the space station's cables and panels becoming underwater rock formations and swaying kelp illuminated by shafts of blue-green light from above, the floating water droplet becoming a small school of tiny fish. Keep their exact pose and the timing of every movement identical."

Before: Astronaut in space station. After: Deep-sea diver

Reference to Video: Subject-Consistent Scenes With Dialogue, Examples from Gemini Omni 1.1 Flash

Give Reference to Video images of a subject, a short reference clip, or both, and it places that subject in a new scene. Scripted dialogue in the prompt comes through with natural, well-timed lip sync, which opens up spokesperson and character-performance use cases. The examples below were made with Gemini Omni 1.1 Flash Reference to Video. Gemini Omni 1.2 Flash Reference to Video takes up to 10 reference images and makes clips of up to 15 seconds.

Reference character speaking a scripted line

The reference photo is a head chef in a white toque. In the generated clip he stands behind the pass counter presenting a plated dish with both hands, looks at camera, and says warmly: "Bon appetit, enjoy every bite," before setting the plate down with a satisfied nod.

Settings: 1 reference image, 16:9, 10 seconds, 1080p. Watch: Open on Scenario

Reference character performing, no dialogue

The reference photo is a jazz vocalist in a sequined gown. In the generated clip she stands at a vintage microphone in a dim smoky club, swaying gently as she sings a held note, sequins catching the amber stage light, a faint upright bass line underneath.

Settings: 1 reference image, 16:9, 10 seconds, 1080p. Watch: Open on Scenario

Extend: Continue Any Video, Examples from Gemini Omni 1.1 Flash Extend

Feed Extend any existing video and it appends 3 to 10 seconds of new footage that continues the same subject, camera style, and native audio, rather than starting a fresh generation. It works from any source clip, including outputs from entirely different models. Both examples below started as clips from Seedance 2.5 and were extended here with Gemini Omni 1.1 Flash Extend. Gemini Omni 1.2 Flash Extend needs a prompt describing how the clip continues, like the ones below, and takes a source video of 30 seconds or less.

Extending a Seedance 2.5 clip: fighter jet

Prompt: "Continue the scene exactly as it ends: the fighter jet continues its steep ascent into the clear blue sky, punching through a thin wisp of cloud as a bright contrail begins to form, then levels off at altitude with sunlight glinting off its fuselage."

Source: Seedance 2.5 clip, extended by 6 seconds. Watch: Open on Scenario

Extending a Seedance 2.5 clip: noir dog with whiskey

Prompt: "Continue the scene exactly as it ends: the dog in the trench coat and fedora lifts the whiskey glass in one paw and takes a slow sip, then sets it back down with a satisfied exhale and leans back into the desk chair, tipping its fedora down with the other paw."

Source: Seedance 2.5 clip, extended by 6 seconds. Watch: Open on Scenario

Gemini Omni 1.2 Flash: Parameters

prompt

What the clip shows, or what to change. Optional on the base model when you give a first frame, and optional on Reference to Video, where your reference subjects appear either way. Required on Edit and on Extend: Extend needs a prompt describing how the video continues, which was optional on 1.1. Upscale takes no prompt. Put dialogue in quotes, as in the ship captain example above.

image

Base model only. An optional first frame the video starts from. Leave it empty to generate from the prompt alone.

lastFrameImage

Base model only. An optional closing frame the clip interpolates toward, for a precise beginning-to-end shot. It requires a first frame and limits the video to 10 seconds.

referenceImages

Images of the subjects to keep consistent. Up to 10 on the base model, where the first and last frames count toward the 10; 1 to 10 on Reference to Video, optional when you give a reference video; and 1 to 10, optional, on Edit, for subject or look consistency. The first frame, last frame and reference images are independent and can all be used in the same generation. The chef and jazz vocalist examples above each used one reference image.

referenceVideo

Reference to Video only. An optional reference clip of 3 seconds or less, used on its own or with up to 10 reference images.

video

The source clip. Edit takes a video of 15 seconds or less. Extend takes a video of 30 seconds or less, from any model, as the two Seedance 2.5 clips above. Upscale takes a video of 15 seconds or less that is 16:9 or 9:16 at exactly 360p, 720p or 1080p (640x360, 1280x720 or 1920x1080, or 360x640, 720x1280 or 1080x1920 vertical).

duration

On the base model and Reference to Video, the clip length: 3 to 15 seconds, 8 by default, and at most 10 with a last frame. On Extend, the seconds of new footage appended after the source: 3 to 10, 6 by default, so an 8-second source extended by 6 seconds gives about 14 seconds. Edit and Upscale keep the source length. The ship captain and chef examples above are 10-second clips, and the two Seedance 2.5 clips were each extended by 6 seconds.

aspectRatio

Base model and Reference to Video: 16:9 (the default) or 9:16. Edit and Extend keep the source's shape. The text-to-video and Reference to Video examples above are all 16:9.

resolution

Base model, Edit, Reference to Video and Extend: 360p, 720p (the default), 1080p or 4K. 360p is a faster draft tier, and 1080p and 4K are upscaled. 360p and 720p cost less on 1.2 than on 1.1. On Upscale, the output resolution must be above the source's: a 360p source goes to 720p or 1080p, a 720p source to 1080p or 4K, and a 1080p source to 4K. Upscale defaults to 1080p. The ship captain and chef examples above were rendered at 1080p.

Cross-Model Pipeline

The Gemini Omni 1.2 Flash models compose. A typical high-impact workflow chains them together, with Nano Banana 2 Lite upstream, Extend to lengthen a shot and Upscale as the finishing step:

1. Nano Banana 2 Lite generates a hero image (character portrait, product hero, or scene still) from text or refs.

2. Gemini Omni 1.2 Flash (base) animates that image into a clip of up to 15 seconds with native audio, using the image as the first frame.

3. Gemini Omni 1.2 Flash Edit produces stylistic or seasonal variants, or a full style and setting reinvention, without regenerating the performance.

4. Gemini Omni 1.2 Flash Reference to Video places the same character in additional scenes using the original hero image as a reference, with spoken dialogue.

5. Gemini Omni 1.2 Flash Extend appends extra seconds onto whichever clip cut a beat short, even one generated by a different model entirely.

6. Gemini Omni 1.2 Flash Upscale raises the clips you keep from a fast draft, with their length and audio unchanged: a 360p draft to 720p or 1080p, a 720p draft to 1080p or 4K. Upscale takes clips of 15 seconds or less, so an extended clip that runs longer cannot go through it: for those, render the Extend step directly at 1080p or 4K instead.

Result: a full multi-scene story with a locked-in cast, dialogue and native audio throughout, generated end-to-end on Scenario.

Tips for Better Results

1. Script the audio in the prompt. Dialogue in quotes, ambient sounds by name, mood cues for score. Omni Flash sings when the audio is scripted and drifts when left to implication.

2. For hero shots, feed a controlled first frame. Text-only can drift in composition. Producing the opening image in Nano Banana 2 Lite or GPT Image 2 first, then handing it to Omni, gives you a locked launch point.

3. For edits, always end with a preservation clause. "Keep the exact motion, timing, and camera path unchanged." Otherwise the model may reinterpret the beat.

4. For R2V, one clean reference beats three noisy ones. A single sharp portrait with the subject clearly visible locks identity better than three references with occlusion or motion blur.

5. Describe camera moves as verbs, not presets. "The camera slowly pushes in" beats "35mm lens, shallow depth of field". Preset language often bakes into the frame as visible text or misfires as style.

6. Multi-shot sequences work if you script them beat by beat. Break a 10-second clip into five 2-second beats in the prompt (0 to 2s, 2 to 4s, and so on) and describe what changes each time. Match-cuts and whip pans are honored when explicitly asked.

7. Non-English dialogue may vary in accent. For non-English lines, name the language and accent explicitly in the prompt, or feed a spoken audio reference through a separate TTS model.

Known Limitations

• 15-second maximum per generation. Clips run 3 to 15 seconds, and at most 10 with a last frame. Use Extend to continue a clip, or concatenate several clips for longer beats.

• Only 16:9 and 9:16. No 1:1, 4:5, or 21:9 native.

• Native audio is one pass. You cannot separately request instrumental-only or dialogue-only. If you need clean stems, add audio in post.

• Every clip has audio. No Gemini Omni model has a switch to turn audio off. For a silent video, mute or remove the track after download, or use a model whose Generate Audio switch you can turn off, such as Kling V3 or Veo 3.1.

• Edit inherits duration and aspect from the source. No override.

• Multi-turn editing is not supported. Each Edit run is one-shot. To iterate, re-run with a revised prompt on the original source.

• Audio input is not accepted. None of the siblings take audio references. The launch demos that used audio as a driver are not exposed on Scenario yet.

• Reference to Video takes only a short reference clip. A reference video must be 3 seconds or less; for longer motion, use Edit on the source clip instead.

Additional Limitations for Extend, Reference to Video and Upscale

• Extend expects a standard aspect ratio. Source clips at 16:9 or 9:16 extend reliably. Unusual source aspect ratios (21:9, 4:3) failed in testing. Re-crop or regenerate the source at 16:9 or 9:16 first.

• R2V's content filter can false-positive on photoreal portraits. A reference image that happens to resemble a real person, even an entirely AI-generated one, can trip a "prominent individuals" rejection. Regenerate the reference with more distinct features rather than retrying the same image.

• Upscale accepts only standard sources. The source must be 16:9 or 9:16, 15 seconds or less, and exactly 640x360, 1280x720 or 1920x1080 (360x640, 720x1280 or 1080x1920 for 9:16), and the output resolution must be above the source's. If the clip is already 16:9 or 9:16 but off-size, resize (scale) it to the nearest of these sizes rather than cropping it. If it has another aspect ratio, crop or pad it to 16:9 or 9:16 first, then scale it; or regenerate it.

Use Cases

• Advertising: product spots with native voice-over, brand vignettes, seasonal ad variants of an approved cut, market-specific restyles, hero social clips.

• Games: in-game cinematics, character reveal trailers, marketing shorts with dialogue, episodic character content, restyled trailer variants (photoreal to anime, day to night).

• Film and animation: pre-vis with sound, animatics with scratch dialogue, mood exploration on live-action plates, style testing on hero shots.

• Marketing: repurposing existing clips for new campaigns without reshooting, spokesperson consistency across market variants, mascot in multiple contexts.

• Education: narrated micro-lessons, historical reconstructions with ambient sound, recurring on-screen host across a lesson series.

• Social media: vertical hero clips with dialogue and ambience baked in, one-pass content workflow.