Generate video from text, images, references, or a start and end frame, then keep editing it in conversation. Official numbers such as the 10-second clip ceiling, extension to a 40-second total, 360p drafting, and 1080p or 4K upscale are labeled separately from the controls this workspace currently exposes.
Text to video · Image to video · Start and end frames · Reference to video · Confirm live duration, resolution, and points before submitting
Generated content will appear here
Four reserved demo slots cover generation, references, frame control, and conversational editing. Reviewed videos, posters, and captions can replace these explicit placeholders later.
Official capability demo
A landscape text-to-video sample for checking visual continuity against dialogue, ambience, and sound effects generated together with the picture.
One continuous shot: a camera glides across a dark studio toward a glass bottle while warm edge light reveals the label. Room tone and a soft glass chime, no dialogue.
Confirm live resolution
A mobile image-to-video sample focused on identity, material detail, and restrained camera motion from a single authorized still.
Preserve the person, clothing, and background from the authorized image. Add a slow turn toward camera, shifting window light, and quiet room ambience. No cuts.
Official capability · Route may differ
A paired sample that generates the continuous footage between an approved opening frame and an approved closing frame.
Start on a wide static frame of an empty gallery, end on a close frame of the sculpture. One continuous camera push, no jump cuts, soft footsteps and room tone.
Official capability · Route may differ
A paired sample showing a targeted object or background change while the rest of the shot is preserved across turns.
Change the background to a rainy street at night. Keep the subject, face, clothing, timing, camera path, and original dialogue. Add subtle rain reflections and distant traffic.
Facts checked 2026-09-11:
What Google documents for the model and what this workspace exposes are two separate lists. Read them side by side before you submit a task.
Google DeepMind also states a first-party difference in prompting style: with Veo you need to share precise instructions, while Gemini Omni is designed to work from a less prescriptive prompt. That is an official statement about the two products, not a benchmark result, and it does not guarantee which route or control set is live here.
Gemini Omni 1.1 Flash is Google DeepMind’s video generation model, published on August 27, 2026 and registered on its official model card with the launch stage Preview. It reads text, still images, and video, writes video output that already contains sound, and can be refined over several turns instead of being regenerated from scratch. The model card lists the model ID as gemini-omni-1.1-flash-preview.
The headline capability is scene extension. Official documentation says the model can now analyse up to 10 seconds of previous footage, where earlier models referenced only the final second, and that you can extend a clip in 10-second increments up to a total cumulative length of 40 seconds. Officially documented output is capped at 10 seconds per clip, offered in 16:9 and 9:16, with 360p and 720p generation and 1080p or 4K upscale. Google records 360p drafting as up to 60 percent faster than 720p by system throughput.
Two boundaries are worth stating plainly, because they are frequently reported incorrectly elsewhere. First, the published ceiling is ten seconds per clip; there is no official three-second minimum, and the three-second figure belongs to the video reference input, which is the material you supply rather than the footage you receive. Second, Google publishes no frame rate for this model on any reachable official page, so this page does not claim one.
Nano Banana Pro reaches the model through its own live route. Available modes, duration, resolution, and the point estimate can differ from the model ceiling, so the workspace display is the source of truth at submission time.
The 1.1 update concentrates on longer scenes, cheaper iteration, and higher delivery resolution.
Continue an existing clip instead of restarting it. Officially documented as 10-second increments up to a cumulative 40-second total.
The model analyses up to 10 seconds of the preceding footage, a documented jump from earlier models that referenced only the final second.
Supply an opening and a closing frame and let the model generate continuous footage between them, which suits orbiting moves, transitions, and looping clips.
Iterate at 360p for storyboard and composition checks before paying for a full resolution pass. Google records up to 60 percent faster generation versus 720p.
Finish approved drafts as 1080p or 4K output for production delivery rather than native generation at those sizes.
Speech, music, and sound effects are generated alongside the picture, so dialogue, ambience, and score can be directed in the same prompt.
Refine a result across turns by describing only what should change, which keeps the rest of the shot stable instead of re-rolling it.
Guide consistency with up to three seconds of reference footage, and up to three videos in a single prompt, alongside up to ten images.
The 1.1 update concentrates on longer scenes, cheaper iteration, and higher delivery resolution.
Prototype a reveal with stated camera direction, authorized references, dialogue, and sound design before committing to a full shoot.
Block out a sequence at 360p, agree on framing and pacing, then re-render the approved takes at higher resolution.
Chain extensions from a single clip toward a documented 40-second ceiling when a story needs more room than one generation allows.
Hand the model an approved opening and closing composition and let it fill the movement between them for trailers, logos, and scene changes.
Test whether authorized image and video references hold identity, wardrobe, props, and lens language across a sequence.
Correct a background, object, or lighting decision in conversation while leaving performance, timing, and camera path untouched.
Google DeepMind’s official guide builds prompts from five elements: shot framing and motion, style, lighting, location, and action. Each template below names the elements it leans on, and the guide explicitly says you do not have to over-describe, because the model works from the overall intention.
Framing and motion plus lighting plus action: one uninterrupted camera move with synchronized sound.
One continuous shot. The camera glides across a dark studio toward a glass fragrance bottle. Warm edge light finds the label as fine mist drifts through the frame, and the bottle itself stays unchanged. Sound design: quiet room tone, one soft glass chime, no dialogue.
Action plus lighting: hold an authorized still stable and describe only the motion and ambience you want added.
Preserve the person, facial identity, hairstyle, clothing, and background from the authorized reference image. The subject turns slowly toward camera as soft window light shifts. Natural fabric movement and quiet room ambience, no scene cuts.
Framing and motion: state the two approved compositions and ask for one unbroken move between them.
Begin on a wide static frame of an empty gallery, end on a close frame of the central sculpture. One continuous camera push between the two, no jump cuts. Soft footsteps and low room tone, cool daylight from the left.
Action plus style: give each authorized reference one clear job so the directions do not compete.
Use the authorized character reference for identity and wardrobe, and the product reference for shape and colour. A controlled tracking shot moves through a minimal gallery. Keep both references consistent, use cinematic lighting, and add quiet footsteps and room ambience.
Framing and motion at draft resolution: prove the shot works before spending on a full pass.
Draft pass for a six-beat sequence. Beat one: wide establishing frame of a harbour at dawn. Each following beat pushes the camera closer to a figure on the dock, ending on a medium close-up. Consistent light direction across all beats, ambient water and gulls throughout.
Action with an explicit preserve list: change one thing and name everything that must not move.
Change the background to a rainy city street at night. Preserve the subject, face, clothing, timing, camera movement, original dialogue, and the foreground objects. Add restrained rain reflections and distant traffic ambience.
Both are preview models from Google, so this is a generation comparison rather than a stable-versus-preview one. Rows come from the two official model cards. On prompting style, Google DeepMind states that Veo needs precise instructions while Omni is designed to work from a less prescriptive prompt.
| Comparison | Gemini Omni 1.1 Flash | Omni Flash (previous) |
|---|---|---|
| Launch stage | Preview · released August 27, 2026 | Preview · released June 30, 2026 |
| Official model card ID | gemini-omni-1.1-flash-preview | gemini-omni-flash-preview |
| Official resolutions | 360p, 720p, 1080p, 4K | 720p |
| Scene extension | Supported, up to 40 seconds total | Not supported |
| Video from an image | Supported | Not supported |
| Video from references | Supported | Not supported |
| Videos from first and last frames | Supported | Not supported |
| Documented retirement | Not announced | June 30, 2027 |
Video tasks use Nano Banana Pro points. The estimate moves with duration, resolution, route, and the live model catalog, so this page does not promise a fixed price. Google publishes no per-second figure for this model, and token pricing is not comparable to a video task.
View plans and point packsCheck the estimated points in the workspace before submitting, then reconcile them against the completed task.
Available resolutions load from the live route. The model card listing 4K does not mean every route always exposes 4K, and delivery at those sizes is an upscale rather than native generation.
Pick text, image, start/end frame, or reference mode, and confirm the model name shown in the workspace.
Use material you have permission to upload, and state what each image or clip should control. Up to ten images, or up to three reference clips of three seconds each, are documented.
Cover shot framing and motion, style, lighting, location, and action. Google’s guide notes that the model works from your overall intention, so heavy detail is optional rather than required.
The guide suggests terms such as one continuous shot or oner, static, locked off, fixed, push in, punch in, dolly zoom, natural smartphone zoom, film camera, and webcam style.
Prove the shot at 360p, extend or edit it as needed, then re-render the approved take and upscale to 1080p or 4K before checking duration, resolution, and points.
This model is registered on its official card with the launch stage Preview, so capabilities, limits, and availability can change. The same card lists global availability rather than regional restrictions, and note that audio is an output modality only: audio is not accepted as an input.
Some numbers are simply not published. Google documents no frame rate for this model and no per-second generation price, so this page does not invent either. Where a figure is quoted here, it comes from the official model card or the release announcement, and both are linked at the bottom of the page.
Generated output carries C2PA Content Credentials for provenance, as recorded on the official model card. Before publishing, review identity, hands, objects, readable text, physics, dialogue, audio, and transitions, and treat any output from a preview model as subject to revision.
Use only inputs you have the right to use. Do not create deceptive impersonation, harassment, privacy violations, copyright infringement, or fabricated evidence. Model access, limits, and platform routes may change.
Model capabilities, stages, and limits come from Google’s model card, release announcement, and DeepMind prompt guide. Platform routes and pricing are labeled separately. Google’s API documentation pages are listed as well; they require a Google sign-in to open.
Open the live workspace, confirm the available route and the point estimate, then turn a prompt, a keyframe pair, or an authorized reference into video.