How Gemini Omni Changes the Way You Write Short-Form Video Prompts

Google launched Gemini Omni at I/O 2026 yesterday, May 19. The first model in the family, Omni Flash, is already rolling out inside the Gemini app, Flow, and YouTube (free for Shorts and Create users), with developer API access promised in the coming weeks. The headline framing from DeepMind is that Omni can "create anything from any input, starting with video," and that framing is doing more work than usual. The real shift is not that Google has a new video model. The real shift is that Omni accepts a mixed prompt of image, audio, video, and text, and reasons across all of it in one pass instead of stitching the pieces together.

For anyone writing prompts for short-form video, this changes the shape of a good prompt. The prompt is no longer a paragraph of text plus maybe one reference image attached for style. The prompt is now a small bundle of evidence that the model treats as one scene.

This post covers what Omni Flash actually does, what kind of prompt the model now expects, and where the model still falls down.

What Gemini Omni actually is

Gemini Omni is a multimodal generative model that produces and edits video from any combination of input types. The first variant, Omni Flash, generates clips capped at ten seconds. Google has been clear that the ten-second cap is a deployment decision rather than a model constraint, which suggests longer outputs are available internally but held back for rollout reasons. A higher-end Omni Pro is planned with no release date.

The model ships with a few specific properties worth naming:

• Multimodal prompts. A single Omni prompt can contain text, one or more images, an audio clip, and a reference video. The model treats these as one input, not as text plus attachments.
• Conversational refinement. Once Omni produces a clip, the next prompt in the conversation can change one element ("redo with the camera lower", "swap the audio for the original reference") and the model edits the same generation. This is a meaningfully different loop than regenerating from scratch.
• SynthID watermark on every clip. Every Omni output carries an imperceptible watermark verifiable through the Gemini app, Chrome, and Google Search.
• Distribution inside Google's surfaces. Omni is in Search, the Gemini app, Flow, and YouTube. It is not a standalone product, which is how it ends up free for YouTube creators and how it competes with standalone video tools on convenience rather than capability alone.

The piece that matters most for prompt writing is the first one. The other three change the workflow around the prompt; multimodal input changes the prompt itself.

Why a multimodal prompt is a different object than a text prompt with attachments

Earlier video models accepted text as the primary input. Reference images were possible in some products, but the model treated them as style hints, not as constraints on the scene. The reference said "match this look." It did not say "the subject in your output is the subject in this image, in this exact pose."

Omni treats every input as part of the scene description. The image is not a style hint. The image is the subject. The audio is not a music suggestion. The audio is the soundtrack the model will sync to. The reference video clip is not a vibe board. The reference video is the camera language the model will copy.

This collapses three or four things that used to be separate prompt elements into one structured input. A 2025-era prompt for a fitness clip looked like this:

"A young woman runs through a neon-lit Tokyo alleyway at night, slow motion, low angle, 35mm lens, energetic synthwave music, dramatic shadows."

An Omni-shaped prompt for the same clip looks like an image of the runner, a five-second snippet of the synthwave track, and a much shorter text prompt: "Same runner, Tokyo alleyway at night, low angle, slow motion, sync motion to the beat."

The second prompt has more information in it, but less of it is in the text. The image carries the subject's face, body type, and outfit. The audio carries the BPM, the energy level, and the moment the drop hits. The text only needs to carry the things the image and audio cannot: the location, the camera angle, the time-domain choices.

What this changes about writing the text portion of the prompt

Three concrete shifts.

The text describes only what the other inputs cannot. If the image already shows the subject, you do not describe the subject in text. If the audio already carries the energy, you do not write "energetic" or "uplifting." Repeating information that the other inputs already encode confuses the model rather than reinforcing the request. The text portion of an Omni prompt is shorter than a text-only prompt for the same shot, often by half.

Camera language goes in the text or in a reference clip, not both. If you attach a reference video that shows a specific camera move, do not also describe that move in the text. Pick one. The reference video is more reliable for complex moves (Hitchcock zoom, whip pan, parallax tracking), and the text is more reliable for simple moves (slow push-in, static, handheld). Use each where it is stronger.

Time-domain choices belong in the text. What happens at second zero versus second nine is something the model cannot infer from a still image. Where the camera starts versus where it ends, what the subject does at the open versus at the close, where the beat hits relative to the cut. This is the highest-value information in the text portion of an Omni prompt. Spend the prompt budget here.

The implication is that the text portion of an Omni prompt has a narrower job than a text-only prompt. It is no longer a description of the scene. It is a list of the things the model cannot read off the image and audio.

Working with the ten-second cap

The ten-second cap on Omni Flash is the most important constraint to design around. Most short-form video for TikTok, Reels, and Shorts runs three to fifteen seconds for the hook, and a ten-second clip covers the hook plus the first beat of the payoff. The cap is not a limit on what short-form creators do; it is roughly the length of the most-watched portion of a Short.

Three practical points about prompting inside the ten-second window:

The prompt should describe one continuous shot. Cuts inside a ten-second clip require either a video reference that already has the cut, or a prompt that the model will probably ignore. Omni Flash is best treated as a single-shot generator, the way 35mm film cameras shoot a single take and then you cut later in the edit.

Pacing matters more than story. Ten seconds is too short for plot. It is long enough for a beat: a setup at second zero, a turn at second four or five, and a resolution at second nine. Writing the prompt as three time markers (zero, mid, end) maps cleanly to how the model lays out the clip.

Audio sync is now part of the prompt, not part of the edit. Because Omni accepts an audio reference and syncs visual motion to it, the beat structure of the audio determines the visual structure of the output. Pick the audio before writing the prompt, not after.

Conversational refinement is a new prompt loop

The conversation loop Omni supports ("redo with the camera lower", "swap the music for the original reference, keep the visuals") is different from regenerating with a tweaked text prompt. The model holds the previous output as state and edits within that state, which produces something closer to a real edit than a re-roll.

This changes how you draft the first prompt. The first prompt does not need to be perfect. It needs to get the subject, the location, and the audio right. The camera angle, the lens, the motion details, the pacing, all of these can be adjusted in follow-up turns. A workflow that used to be "spend twenty minutes writing one prompt, run it, accept the output" becomes "spend two minutes on a rough first prompt, run it, then spend ten minutes refining specific elements through the conversation."

The catch is that conversational refinement uses one credit per turn just like a fresh generation. The economy of the new workflow is "more cheap, small iterations" rather than "one careful iteration." Both produce comparable end results. The conversational path tends to produce more interesting results because each refinement has the previous frame as a reference, not just a text description.

Tradeoffs and limits

Omni is not without rough edges, and a few of them matter for short-form work.

Omni Pro is not out yet. Flash is real and shipping. Pro is announced with no date. If you are betting on a workflow that needs the quality ceiling of Pro, you are betting on something Google has not delivered.

The developer API is "coming weeks" away. Inside Gemini, Flow, and YouTube, Omni is usable today. Outside those surfaces, building Omni into a pipeline requires waiting. For ReezoAI, this means our tools will continue to generate prompts that creators take into Omni manually, rather than calling Omni directly, until the API ships.

SynthID is imperceptible and unavoidable. For most short-form creators this does not matter. For anyone whose downstream platforms strip metadata or care about provenance signals, this is a constraint.

The model has Google's house style. Omni outputs trend toward a clean, well-lit, color-graded look. If you want something rougher (handheld, low-light, blown highlights, intentional grain), specify it loudly and provide a reference clip that demonstrates it. The default will not look like a phone-shot vlog without help.

When to use a different tool

The shift to Omni does not mean every short-form workflow runs through it.

For B-roll and stock-style footage where the input is purely text, a text-to-video model with a longer track record (Veo 3, Sora) is still the right pick. Omni's strength is in the multimodal prompt; if the prompt is only text, the multimodal advantage goes away.

For prompts that need ten or more seconds of continuous output, Omni Flash is the wrong tool today. The cap is real and the model does not generate cleanly past it. Wait for Omni Pro or use a longer-form generator.

For prompts where the reference image is the actual subject (a specific person, a specific product), confirm consent, likeness rights, and brand permissions before generating. Omni is good enough at subject preservation that the output will look like the source, which is a reason to be careful with what you put in, not a feature to lean on.

How ReezoAI's tools fit the Omni workflow

The Reeprompt tool writes the text portion of an Omni prompt. With Omni in the picture, the text portion is shorter and more time-specific than it used to be. Reeprompt's platform-specific hooks (the first-three-seconds pattern for Shorts, the pattern interrupt for TikTok) become more important, not less, because they are exactly the time-domain instructions Omni needs in the text portion.

The PromptForge tool produces JSON-shaped prompts in the Veo schema. JSON prompts remain the right structure for text-to-video workflows that do not have an image or audio reference to lean on. PromptForge's conversation memory (prompt chains, generation tracking, branching) is also a good fit for the Omni-style refinement loop, where each turn edits a small part of the previous prompt rather than rewriting the whole thing.

The biggest shift in the workflow is not in the tool. It is in what you bring to the prompt. A creator who shows up with a reference image, a reference audio clip, and a clear time-domain plan gets dramatically better results from Omni than a creator who shows up with a wall of descriptive text. Spend the time on the reference materials.

Short summary

Gemini Omni does not just generate video. It generates video from a mixed prompt that includes images, audio, and reference clips alongside text, and it treats all of those inputs as part of the scene rather than as style hints. The text portion of a good Omni prompt is shorter than the text portion of a 2025-era prompt, and it carries different information: time-domain choices, camera language the reference clip does not show, and the things only words can express.

The ten-second cap matches the shape of short-form video. The conversational refinement loop changes the economy of iteration. The free distribution inside YouTube Shorts and Create will put this model in front of more creators than any standalone video tool ever has. The prompt-writing habits formed against text-only video models will hold most creators back from getting Omni's actual ceiling, and the gap between a text-only prompt and a multimodal prompt is where the difference shows up.