Updated for Omni 1.2 — November 2026

Google Omni 1.2 Prompt Generator

Generate optimized prompts for Gemini Omni 1.2 Flash — Google's any-to-any video model that turns text, images and video into high-resolution clips with native synchronized audio. Built on Google's official prompting documentation, with 1.2's 15-second clips, natural-language references, scene extension, first/last-frame control and conversational editing baked in.

Describe the main visual scene — subjects, environment, mood, and key visual details

Generated Prompt

Fill in the form and click "Generate" to create an optimized Gemini Omni 1.2 Flash video prompt.

Tip: Describe the motion and temporal progression of your scene. Think in terms of "what happens over time" rather than a static description.

Gemini Omni 1.2 Flash Tips

  • • Quotation marks are for on-screen text, not dialogue. Write speech unquoted after a colon — "A woman says: My name is Clara." Quoting it burns the line into the frame as text.
  • • Omni multi-shots by default. If you want one unbroken take you have to say so: "In a single continuous shot" or "No scene cuts."
  • • Less is more — 30 to 80 words is the sweet spot. Google says Omni needs less prescription than Veo; past ~100 words you start fighting the model.
  • • Skip pixel-level camera specs. ISO, f-stop and mm focal lengths actively degrade coherence. "Shallow depth of field" works; "85mm at f/1.4" does not.
  • • Always state audio intent — even if it is "natural ambient sound only, no music, no dialogue." Explicit negation stopped unwanted stock music in about half of documented tests.
  • • For edits, one instruction per turn, then the exact phrase "Keep everything else the same." Drift starts around turn 5.
  • • 16:9 and 9:16 only — there is no 1:1 on Omni. Clips run 3-15s each on 1.2, extendable to 30s total.
  • • Skip tags like <IMAGE_REF_0> — Google deprecated them in 1.2. Say "the woman in the first reference image" instead.
  • • Draft at 360p, finish at 4K. Same prompt, one tenth the cost while you explore variations.

Gemini Omni 1.2 Prompt Templates

Copy-ready shot briefs built around what this model actually rewards — including Google's own published prompts. Swap the [BRACKETED] parts for your own scene, then paste straight into the tool.

Whip-Pan Reveal (Google 1.1 demo)

Google Official

Google's own 1.1 launch prompt — a designed camera move with two subjects.

A close-up low-angle shot of [SUBJECT A, e.g. a stylish drummer in a beige suit playing a red drum kit] in [LOCATION, e.g. a grand hall] transitions as the camera whip-pans to the side, revealing [SUBJECT B, e.g. an older saxophonist] alongside [SUBJECT C, e.g. a ballet dancer spinning in a white outfit] under [LIGHTING, e.g. soft purple stage lights].

16:9, 10 seconds, natural performance audio, no dialogue.

Claymation Knowledge Explainer

Google Official

Google's flagship world-knowledge prompt. The word "accurate" does the real work.

claymation explainer of [SCIENTIFIC PROCESS, e.g. protein folding], everything is made out of clay, no hands, stop motion, accurate

Sound design: soft tabletop foley only. No music, no dialogue. In a single continuous shot.

Kinetic Typography Hook

Text on Screen

Quoted glyphs render on screen. Note there is no dialogue here at all.

One word on the screen at a time: "[WORD 1], [WORD 2], [WORD 3], [WORD 4], [WORD 5]" Each word appears for 1s with a different animated style, perfect pacing to a rhythm.

Background: [CLEAN BACKDROP, e.g. a matte black studio]. Audio: soft electronic clicks timed to each word. No dialogue. No music.

9:16 vertical, 8 seconds.

Diegetic In-World Text

Text on Screen

Text that lives inside the scene — signs, storefronts, plates. Quote the exact glyphs.

[SCENE DESCRIPTION, e.g. A slow push down a rain-slicked street at dusk].

There is a street sign that says: "[SIGN TEXT]", there is a storefront that says: "[STOREFRONT TEXT]", there's a [OBJECT] with the [LABEL TYPE]: "[LABEL TEXT]".

Keep all on-screen text correct and readable. Sound design: [AMBIENT SOUND]. No dialogue.

Dialogue — Unquoted, After a Colon

Audio

The rule most guides get wrong. Quotes render text; a colon speaks the line.

[SPEAKER DESCRIPTION, e.g. A woman in her thirties wearing a grey overcoat] in [SETTING], [FRAMING, e.g. medium shot, eye level].

In a voice that is [VOICE QUALITIES, e.g. crisp and clear, with a thoughtful, analytical tone and a standard American accent], [SPEAKER NAME] says: [LINE OF DIALOGUE, UNQUOTED]

Sound design: [AMBIENT BED] kept low under the voice. No music. In a single continuous shot.

Surgical Conversational Edit

Editing

Google's documented pattern — short instruction plus one exact preserve phrase.

[ONE CHANGE, e.g. Make the phone invisible. / Change the car color to metallic blue. / Put a fashionable hat on this person.] Keep everything else the same.

Event-Conditional Edit (two turns)

Editing

Turn 1 establishes timed beats, turn 2 edits relative to them. A documented power move.

Turn 1: [SCENE, e.g. A person stands in front of a mirror]. At 0:05, [TRIGGER ACTION, e.g. she touches the mirror] then she does it again at 0:07, 0:08 and 0:09.

Turn 2: [WHAT CHANGES, e.g. The style of the video changes] every time [TRIGGER ACTION, e.g. the person touches the mirror].

Single-Variable A/B Chain

Editing

One variable per turn gives you a clean matched set instead of unrelated renders.

Turn 1: [BASE SCENE, e.g. A person opening a sleek black product box on a wooden desk in a minimalist apartment]. [LIGHTING, e.g. Morning light]. [CAMERA, e.g. Slow zoom in]. In a single continuous shot.

Turn 2: Make it [VARIABLE 1 CHANGED, e.g. sunset light instead of morning]. Keep everything else the same.

Turn 3: Now make [VARIABLE 2 CHANGED, e.g. the box white instead of black]. Keep everything else the same.

Scene Extension

Extension

Each pass appends 3-10s, up to 30s total on Omni 1.2. Timecodes reset to the new portion.

Continue the video. [WHAT HAPPENS NEXT, e.g. The camera slightly pans and we now see she is talking to a man with curly hair, we see the man's back].

[SPEAKER] says: [LINE, UNQUOTED, IF ANY]

Audio: [HOW THE SOUND CHANGES, e.g. the music continues into the chorus]. Keep the character, lighting and location consistent with the previous shot.

Extension — Cinematic Dolly Zoom

Extension

Google's 1.1 demo. The move that used to break at the extension seam.

Continue the video. Execute a cinematic optical dolly-zoom shot. The camera dollies forward while simultaneously zooming out, keeping [SUBJECT]'s [EXPRESSION/POSE, e.g. frozen shocked face] locked at the exact same size.

Audio: [SCORE OR AMBIENCE, e.g. a rising dramatic score under room tone]. No dialogue.

First + Last Frame Interpolation

References

You own both ends; Omni invents only the middle. Same image twice = seamless loop.

A smooth cinematic transition from [WHAT THE FIRST FRAME SHOWS] to [WHAT THE LAST FRAME SHOWS]. In between, the camera [EXACT TRANSITION PATH, e.g. pushes forward and arcs gently right]. No jump cuts.

Use the first image as the opening frame and the second image as the final frame. [DURATION, 3-10 SECONDS], [ASPECT RATIO], [AUDIO INTENT].

Multi-Reference Sequence with Timecodes

References

Six references, the 1.2 maximum. Name each image by order in plain words; the old tags are deprecated.

[0-3s] [SCENE TYPE, e.g. A studio fashion sequence]. Starting with [SUBJECT] from the first image, [ACTION] the item from the second image
[3-6s] Then we see [SUBJECT] from the third image [ACTION] the item from the fourth image
[6-10s] And finally [SUBJECT] from the fifth image [ACTION] the item from the sixth image

Use the given images as references for video generation. The images should not be used as literal initial frames.

Role-Assigned Multi-Modal Mix

References

Label what every reference is FOR — unlabelled references get blended together.

The first image is [IDENTITY], the second image is [PRODUCT], and the reference video sets [CAMERA MOVE].

[SUBJECT] from the first image [ACTION] holding the [PRODUCT] from the second image, with the camera language of the reference video.

Use the given image(s) as references for video generation. The images should not be used as literal initial frames. Use the given video(s) as references. Do not use them as a source for video editing.

Motion Transfer — Identity from Image

References

Video reference carries motion, image reference carries identity. Separates performance from casting.

Apply the pose and motion from the reference video to the character in the reference image.

Place them in [ENVIRONMENT]. Keep the character's face, hair and wardrobe exactly as the reference image establishes. [CAMERA]. Audio: [SOUND].

Use the given video(s) as references. Do not use them as a source for video editing.

Sketch-to-Video

Google Official

The final clause is what stops your drawing appearing in the output.

turn this into [REALISM LEVEL, e.g. realistic footage], using the drawing only as a guide for movement, do not show the drawing in the final video

[SUBJECT AND SETTING DETAIL]. [LIGHTING]. Sound design: [AUDIO]. No dialogue.

Single Continuous Shot (oner)

Cinematic

Omni multi-shots by default — continuity is opt-in and needs a literal phrase.

[SCENE AND SUBJECT, e.g. A young product designer sits at a small desk beside a rainy window and opens a sketchbook].

[ONE OR TWO ACTIONS MAXIMUM]. The camera [CAMERA MOVE, e.g. starts close on the pencil tip, pulls back to a medium shot, then orbits gently left].

[LIGHTING, e.g. Warm desk lamp light, cool blue rain outside, shallow depth of field]. In a single continuous shot. No scene cuts.

Sound design: [AMBIENCE]. No music, no dialogue. 16:9, 10 seconds.

Multi-Shot Sequence with Timecodes

Cinematic

When you actually want cuts, author the beats explicitly rather than leaving it to chance.

[0-3s] [BEAT ONE]
[3-6s] [BEAT TWO]
[6-10s] [BEAT THREE]

Consistent throughout: [WHAT MUST NOT CHANGE, e.g. the same character, the same warm tungsten lighting, the same location]. [STYLE].

Sound design: [AUDIO ACROSS ALL BEATS]. 16:9, 10 seconds.

Physics-Grounded Demonstration

Explainer

Omni simulates momentum, deformation and friction rather than animating an approximation.

[OBJECT] [PHYSICAL ACTION, e.g. bouncing on a hardwood court]. Each [EVENT] has realistic physics: [SPECIFY — e.g. ball deformation on impact, friction slowing the bounces naturally].

[SPEED, e.g. real time, not slow motion]. Camera: [ANGLE AND MOVE]. Lighting: [LIGHTING].

Sound design: realistic impact audio synchronised to the moment of contact. No music. In a single continuous shot.

Explainer with Native Narration

Explainer

Narration is generated in the same pass and timed to the animation — no separate VO step.

A [STYLE, e.g. skeuomorphism stop motion] explainer about [TOPIC, e.g. how the brain hippocampus works] with a compelling voiceover.

Accurately visualize [WHAT MUST BE CORRECT]. Camera: [FRAMING]. Lighting: [LIGHTING].

A narrator says: [SHORT SCRIPT, UNQUOTED, SIZED TO THE CLIP LENGTH]

Quiet ambient pad underneath. No sound effects. 16:9, 10 seconds.

Product Hero from One Photo

Product & Ads

Image-to-video keeps the real SKU. Prompt motion only — do not re-describe the photo.

Use the attached product photo as the starting frame. The [PRODUCT] rotates slowly on [SURFACE, e.g. a matte studio pedestal], [LIGHTING, e.g. soft key light from the left], shallow depth of field. Keep the logo sharp and readable throughout the rotation.

In a single continuous shot. Sound design: [AUDIO, e.g. calm ambient bed]. No dialogue. [ASPECT RATIO], [DURATION].

Audience Variant from One Master

Product & Ads

Conversational state holds the product constant while pace and palette are re-targeted.

Turn 1: Generate a [DURATION] product ad for [PRODUCT]. [TREATMENT, e.g. Cinematic close-ups, premium feel, deep blacks]. [AUDIO MOOD]. End on the product alone, centered.

Turn 2: Now generate a variant of the same ad targeted at [AUDIENCE]. [NEW TREATMENT, e.g. Faster cuts, brighter colors, social-feed energy]. Keep the product exactly the same.

UGC-Style Creator Ad

Product & Ads

Handheld realism plus lip-synced delivery — the format that reads as genuine.

[DURATION] clip in casual handheld [DEVICE] vlog style. [PERSON DESCRIPTION] walks down [LOCATION] holding the camera in selfie mode, talking energetically about [TOPIC].

Natural skin texture, no filter, slight motion blur. Background: [BACKGROUND].

She says: [SHORT SCRIPT, UNQUOTED — keep it to what fits the clip length]

Sound design: [STREET AMBIENCE] under the voice. No music. 9:16 vertical.

Restyle While Preserving Motion

Editing

Style transfer framed as a verb on the existing scene. Use one style reference only.

Take the attached clip and restyle it as [TARGET STYLE, e.g. a hand-painted animated scene]. Preserve the camera motion and timing exactly. Replace [WHAT CHANGES, e.g. the modern setting with a 1980s countryside]. Keep the subject's posture and expression coherent across frames.

Minimal version: Remake this video in [STYLE] aesthetic. Keep everything else the same.

Vertical Reel from Reference Stills

Social

Multi-image reference preserves the real room, signage and products.

The first image is [WIDE SHOT], the second image is [DETAIL], and the third image is [HERO ITEM].

A [DURATION] vertical 9:16 clip set in [PLACE] from the reference images. Handheld, [LIGHTING, e.g. natural morning light]. The camera drifts from [START POINT], past [MIDPOINT], and settles on [END POINT].

Keep the interior, signage and products exactly as in the references. In a single continuous shot.

Sound design: [SPECIFIC AMBIENT SOUNDS]. No music, no dialogue.

Trigger-Anchored ASMR Sound Design

Audio

Attaching a sound to a specific visual event beats vague audio direction measurably.

[SCENE, e.g. A hand moves slowly through a wall of ferns in a dim greenhouse].

Add [SOUND, e.g. soft harp notes] synchronized to exactly when [TRIGGER EVENT, e.g. the hand touches each fern leaf].

Underneath: [AMBIENT BED, e.g. faint dripping water and distant birdsong]. No music, no dialogue, no narration.

Camera: [FRAMING]. In a single continuous shot. [ASPECT RATIO], [DURATION].

Interior Walkthrough from a Photo

Cinematic

Real-estate and architecture — the camera arc Google's own Home Tour demo showcased.

A single continuous [DURATION] interior walkthrough starting from this photograph. The camera pushes slowly forward through [ROOM], arcs [DIRECTION] past [FEATURE], and pulls back to reveal [FINAL REVEAL].

[LIGHTING, e.g. Late-afternoon golden-hour light]. No people. Keep the furniture, flooring and fixtures exactly as photographed.

Use this image as the starting frame. Sound design: quiet room tone only. No music, no dialogue. No scene cuts.

What's new in Omni 1.2

Generally available November 5, 2026, across the Gemini API, Google Cloud, Flow, Google Vids and Google Ads. Omni 1.1 API traffic routes to 1.2 automatically from December 3, 2026.

15-second native clips

The headline change. Each generation now runs 3 to 15 seconds in one-second steps (default 10), up from 1.1's 10-second ceiling. First and last frame interpolation stays at 3 to 10 seconds.

About 30% cheaper

Google cut the price roughly 30% versus Omni 1.1, to about $0.07 per second of 720p video. A 10-second 720p clip now costs around $0.70.

Editing on full 15-second clips

Multi-turn conversational editing works on the whole 15-second clip: swap a product or background, change lighting or weather, or switch aspect ratio through follow-up turns instead of regenerating.

Better reference following

Google cites improved character and product consistency from reference inputs. Up to 6 reference images per request (3 per character, 2 characters), plus up to 3 three-second video references on clips of 10 seconds or less, 1 on longer clips.

Tags are deprecated

The <FIRST_FRAME>, <LAST_FRAME>, <IMAGE_REF_N> and <VIDEO_REF_N> control strings are deprecated and "unsupported in future models." Describe how each image or video should be used in plain language. The generator now writes prompts this way.

Extension: 30 seconds total

Extension still appends 3 to 10 seconds per pass from the final 10 seconds of context, now from a 15-second base. Google's 1.2 docs put the total at 30 seconds. Append-only, and the timecode origin resets in each extension.

Multi-speaker dialogue

Google cites more natural dialogue, clearer lip-sync and distinct multi-speaker conditioning. The generator now names each speaker and gives each one their own voice description.

Longer continuous moves

Google says 1.2 holds continuous camera moves such as FPV flythroughs, zoom-outs and orbits with fewer jump cuts, which suits 15-second single takes. Both moves are now in the camera list.

What is Gemini Omni?

Gemini Omni is Google DeepMind's any-to-any generative video family, announced at Google I/O on May 19, 2026 under the banner "create anything from anything." It collapses a previously fragmented multimodal stack — text-to-video, image-to-video, video editing, audio generation — into a single model with a single editing surface, combining an intuitive grasp of physics (gravity, kinetic energy, fluid dynamics) with Gemini's knowledge of history, science and cultural context.

Gemini Omni 1.2 Flash (gemini-omni-1.2-flash) became generally available on November 5, 2026, replacing Omni 1.1 Flash (August 27, 2026). It runs in Google Flow, Google Vids, the Gemini app (rolling out to Google AI subscribers, 15-second clips on Ultra), Google AI Studio, the Gemini API, Gemini Enterprise Agent Platform and Google Ads. On the Artificial Analysis video arena, Omni Flash ranks among the top text-to-video and image-to-video models (rankings shift week to week), competitive rather than dominant.

A note on accuracy: much of the third-party writing about this model repeats claims Google has never made — a "Gemini Omni Pro" variant, specific audio codecs, per-generation credit costs. None of those have a primary source. Everything on this page is drawn from Google's own documentation, model card and pricing pages.

Core Capabilities

Any-to-Any Input

Mix text, images and video references in one prompt — Omni reasons across all of them. Audio reference upload is not yet supported in the API.

Conversational Editing

Stateful multi-turn edits. The model remembers the video and applies changes while preserving what you did not mention. Reliable to about 4 turns.

World-Knowledge Grounding

Accurate explainers for technical topics — physics, biology, history, chemistry, architecture — from a named metaphor rather than exhaustive description.

Physics Reasoning

Gravity, kinetic energy, fluid dynamics and cloth simulation are modelled, not approximated. Internal reasoning plans motion before frames are synthesized.

Native Synchronized Audio

Ambient sound, SFX, dialogue and music generated in the same forward pass as the video — not dubbed on afterwards.

Style Transfer

Apply claymation, anime, watercolour, risograph or chalk-on-blackboard while preserving the original motion. One style reference at a time.

SynthID + C2PA

Every output carries an imperceptible pixel and audio watermark plus C2PA Content Credentials, verifiable in the Gemini app, Chrome and Google Search.

Avatar Mode

Create videos featuring a digital version of yourself with your own voice. Enrolment requires explicit consent and is 18+; a single selfie is not enough.

Omni 1.2 Flash Specs

Straight from Google's API documentation and pricing pages. These are the constraints the generator writes against.

Model IDgemini-omni-1.2-flash (1.1 requests route here from Dec 3, 2026)
ReleasedNovember 5, 2026 (GA)
Duration3–15s per generation (3–10s for first/last frame); extension adds 3–10s per pass, up to 30s total
Frame rate24 FPS
Aspect ratios16:9 (default) and 9:16 only — no 1:1
Resolutions360p · 720p (default) · 1080p (upscaled) · 4K (upscaled)
Task modestext_to_video · image_to_video · reference_to_video · edit · extend
Reference tagsDeprecated in 1.2 — describe each reference in natural language
Image referencesUp to 6 (max 3 per character, 2 characters)
Video referencesMax 3 clips × 3s for 3–10s outputs, 1 clip for 11–15s; new generations only; reference audio ignored
Audio reference uploadNot supported in the current API version
Unsupported paramsNo negative prompt, no seed, no temperature, no system instructions
API pricing~$0.07 per second at 720p (about 30% below Omni 1.1); no free API tier
Free accessGoogle Vids (15s clips for eligible accounts); Google Flow free tier; YouTube Shorts & Create
ProvenanceSynthID watermark + C2PA Content Credentials on every output

The Official Omni Prompting Framework

Drawn from Google DeepMind's Omni prompt guide, the Gemini API documentation and Google's video best-practices pages — plus the measured findings from published hands-on testing.

1

Quotes render text; colons speak

The highest-impact rule on Omni, and the one nearly every guide gets wrong. Google: "To prevent the model from rendering text in the video, use a colon (:) after the speaker's action to denote speech and avoid using quotation marks." Write "A woman says: My name is Clara." for audio. Save double quotes for glyphs you want in frame.

2

Less prescriptive than Veo, on purpose

DeepMind states the divergence outright: "With Veo, you need to share precise instructions to get the best results. But with Gemini Omni, you don't have to be as prescriptive." Cover the five elements — framing and motion, style, lighting, location, action — in 30 to 80 words and let world knowledge fill the rest. Ported Veo prompts generally need trimming by 40 to 60%.

3

Single shot is opt-in

Omni builds a short multi-shot narrative by default. If you want one unbroken take you must ask: "In a single unbroken scene," "In a single continuous shot," or "No scene cuts." For deliberate sequences, author beats with timecodes — "[0-3s] ... [3-6s] ..." — but never chain three distinct events in one short clip; Google documents that as producing muddled output.

4

Name each reference's role in plain words

Google deprecated the old control strings (<FIRST_FRAME>, <IMAGE_REF_0>, <VIDEO_REF_0>) in 1.2: "Describe how each image or video should be used in natural language instead." Open with one sentence that gives every reference a job — "the first image is the character, the second is the product, the reference video sets the camera move" — because unlabelled references get blended. Up to 6 images per request. In image-to-video, prompt motion only; re-describing the image degrades results.

5

Edits are surgical, and there is one preserve phrase

Google: "Simple prompts work best for video editing. Overly descriptive prompts can lead to unintended changes." The documented idiom is exactly "Keep everything else the same." One instruction per turn. Their own example: not a 40-word removal request, just "Make the phone invisible. Keep everything else the same." Reliable ceiling is around 4 turns before drift compounds.

6

World knowledge is the moat

For explainers, name the concrete visual metaphor and let Gemini supply accurate detail. Google's flagship prompt is six words of style plus one crucial word: "claymation explainer of protein folding, everything is made out of clay, no hands, stop motion, accurate." Use cultural touchstones, historical eras and scientific terms directly instead of granular description.

7

Audio is generated, not dubbed

Always state audio intent — even when the answer is "natural ambient sound only, no music, no dialogue." Google's own examples use a "Sound design:" prefix. Anchor sound to visual events where you can ("harp notes synchronized to each time she touches a leaf"); trigger-anchored audio measurably beats vague direction, and explicit negation stopped unwanted stock music in roughly half of documented tests.

8

Draft at 360p, finish at 4K

Google's own launch post for 1.2 describes the loop: draft at 360p or 720p, then upscale the winners to 1080p or 4K. Run 3 or 4 variations changing one variable each, compare, then render the winner at high resolution — the cheap draft tier keeps the exploration phase affordable. Google also suggests fixing a single detail with a follow-up edit instead of regenerating the whole clip.

Prompts Published by Google

Verbatim prompts from Google's own launch posts, DeepMind pages and API documentation — the clearest signal of what this model expects.

Whip-pan reveal — 1.1 launch demo
“A close-up low-angle shot of a stylish drummer in a beige suit playing a red drum kit in a grand hall transitions as the camera whip-pans to the side, revealing an older saxophonist playing alongside a ballet dancer spinning in a white outfit under soft purple stage lights.”
Scene extension — dolly zoom on a held reaction
“Continue the video. Execute a cinematic optical dolly-zoom shot. The camera dollies forward while simultaneously zooming out, keeping the character’s frozen shocked face locked at the exact same size.”
World-knowledge explainer — Google’s flagship prompt
“claymation explainer of protein folding, everything is made out of clay, no hands, stop motion, accurate”
Kinetic typography — note the quoted glyphs
“One word on the screen at a time: "did, you, know, that, Omni, can, do, awesome, text?" Each word appears for 1s with a different animated style. No dialogue.”
Surgical edit — the documented simplification
“Make the phone invisible. Keep everything else the same.”
Extension with a reference image — 1.2 docs
“Extend this video: have the character shown in the reference image enter the scene and wave.”
First + last frame — 1.2 docs (no tags)
“A smooth cinematic transition from a lush green forest at sunrise to a snowy forest under a starry night sky.”
Sketch-to-video — the crucial final clause
“turn this into realistic footage, using the drawing only as a guide for movement, do not show the drawing in the final video”
Motion transfer — 3 video references into one space
“Use the three uploaded videos of dancers and replace them with the provided characters. Have them perform their individual dances from the reference videos, all together in the large, open space from the provided image.”

Eight Documented Omni Mistakes

Every one of these is named in Google's own documentation or reproduced across published hands-on testing.

Avoid

A woman says: "My name is Clara."

Instead

A woman says: My name is Clara.

Quotation marks tell Omni to render glyphs on screen. Quoting dialogue is why lines end up burned into the frame instead of spoken.

Avoid

Please remove the cell phone that the person is holding in their hand and fill in the background so it looks like they are just holding their hand empty.

Instead

Make the phone invisible. Keep everything else the same.

Google’s own before/after. Overly descriptive edit prompts confuse the diffing engine and cause unintended changes.

Avoid

Shot on 85mm at f/1.4, ISO 400, anamorphic.

Instead

Shallow depth of field, warm practical lighting.

Pixel-level camera specs actively degrade coherence on Omni. Cinematographic facts steer it; technical metadata does not.

Avoid

Assuming you get one unbroken take.

Instead

In a single continuous shot. No scene cuts.

Omni builds a short multi-shot narrative by default. Continuity is opt-in, and there are three literal phrasings that trigger it.

Avoid

A negative prompt field with "wall, frame, blurry".

Instead

No text overlay on screen. No extra sound effects.

Omni has no negative-prompt parameter. Negatives go inside the prompt as instructions — the exact opposite of Veo’s convention.

Avoid

She walks in, then sits down, then the lights change, then she stands.

Instead

She walks in and sits down as the lights dim.

Chaining three or more distinct events in one short clip is documented to produce muddled or incomplete video. One or two actions per 10 seconds.

Avoid

Re-describing everything already visible in your reference image.

Instead

She turns toward the window and smiles.

In image-to-video, redundant description confuses the model. Prompt motion only and refer to the subject generically.

Avoid

A 1:1 square clip for Instagram.

Instead

9:16 vertical, or 16:9 landscape.

Omni supports 16:9 and 9:16 only. There is no square output — crop in post if you need one.

Frequently Asked Questions

What is Gemini Omni 1.2 Flash?

Gemini Omni 1.2 Flash (model ID gemini-omni-1.2-flash) is Google's any-to-any generative video model, generally available since November 5, 2026. It accepts text, images and video freely mixed in a single prompt and outputs video with native synchronized audio generated in the same pass. It replaces Omni 1.1 Flash (August 27, 2026): API requests to gemini-omni-1.1-flash route to 1.2 automatically from December 3, 2026. The original Gemini Omni was announced at Google I/O on May 19, 2026.

What is new in Omni 1.2 versus Omni 1.1?

Four things. (1) Longer clips — 3 to 15 seconds per generation in one-second steps, up from 10 seconds. (2) A price cut of about 30%, to roughly $0.07 per second at 720p. (3) Multi-turn conversational editing on full 15-second clips, so you can swap a product or background, change lighting or weather, or switch aspect ratio without regenerating. (4) Better reference following, with Google citing improved character and product consistency. Google also deprecated the old control-string tags (<FIRST_FRAME>, <IMAGE_REF_0> and so on) — describe how each reference should be used in plain language instead.

How long can Gemini Omni 1.2 videos be?

Each generation produces 3 to 15 seconds (10 by default), at 24 FPS. First and last frame interpolation is limited to 3 to 10 seconds. From there you extend in further passes of 3 to 10 seconds each, up to 30 seconds total, with the model reading the final 10 seconds as context. Extension is append-only — you cannot prepend or insert mid-clip — and the timecode origin resets, so "after 2s" in an extension prompt means 2 seconds into the new portion, not the original clip.

Should dialogue go in quotation marks?

No — and this is the rule most prompt guides get backwards. Google's best-practices documentation says to use a colon after the speaker and avoid quotation marks, because quotes cause the model to render the line as on-screen text instead of speaking it. Write "A woman says: My name is Clara." unquoted for audio. Reserve double quotation marks for text you actually want rendered in frame, like a street sign or a kinetic-typography sequence.

What aspect ratios and resolutions does Omni 1.2 support?

16:9 (default) and 9:16 only — there is no 1:1, 4:3 or 21:9 option. Resolutions are 360p, 720p (default), 1080p and 4K, with the caveat that Google's own API docs label 1080p and 4K as "upscaled" rather than natively generated.

Does Gemini Omni generate audio, and can I upload an audio reference?

Omni generates synchronized native audio — ambient sound, SFX, dialogue and music — in the same forward pass as the video. Google's 1.2 launch post mentions audio cues and short voice references, but the API documentation says uploading audio inputs or audio references is unsupported in the current version, so in the API describe the audio you want in words instead. 1.2 also adds distinct multi-speaker conditioning: give each speaker a name and their own voice description. English is the only language Google lists as fully supported. Voice editing is not supported anywhere, and you cannot add dialogue when extending an uploaded video of someone talking.

How do negative prompts work on Omni?

There is no negative-prompt parameter at all — nor seed, temperature, or system instructions. Google's guidance is to put negatives directly in the prompt as instructions: "No dialogue," "No text overlay on screen," "No scene cuts." This is the opposite of Veo, where a separate negativePrompt field takes bare noun lists and the word "no" is discouraged. Mixing the two conventions up is the most common Omni prompting mistake.

What does Omni 1.2 cost?

API billing is token-based. Google prices Omni 1.2 at about $0.07 per second of 720p video, roughly 30% below Omni 1.1's $0.10 per second, so a 10-second 720p clip costs about $0.70. 360p drafts cost less and the 1080p and 4K upscales cost more, which is why the draft-low, finish-high workflow is worth building around. There is no free API tier; Google Flow and the YouTube Shorts / YouTube Create apps have offered free Omni access, and Google Vids includes 15-second Omni 1.2 clips for eligible Google account and Workspace users at no additional cost.

Where can I use the generated prompts?

Paste them into Google Flow, Google Vids, the Gemini app (1.2 is rolling out to all Google AI subscribers, with 15-second clips for Ultra), YouTube Shorts, the YouTube Create app, Google AI Studio, the Gemini API, Gemini Enterprise Agent Platform or Google Ads. Note that editing or extending uploaded video is unavailable in the EEA, Switzerland and the UK — model-generated clips can be edited everywhere.

Is this tool free?

Yes! You get 1 free Gemini Omni prompt generation per day. For unlimited generations across all 30+ AI tools, sign up for a Promptslove membership.

Want Unlimited Omni Prompt Generation?

Get unlimited access to the Google Omni Prompt Generator, all 30+ AI tools, 29,000+ premium prompts, courses, and resources.