Gemini Omni just shipped — chat-edit your avatar
Google's new Gemini Omni Flash lets you generate and edit short videos by chatting, and it ships with a personal avatar mode. Here is what changes for your reels this week.
Google shipped Gemini Omni Flash on May 19, 2026, the first video model in the new Omni family. It generates and edits short video from any mix of image, audio, video, and text inputs.
The launch includes a personal avatar mode. You record your own voice and likeness once, then build videos that look and sound like you, all watermarked invisibly with SynthID by default.
Flash-tier clips cap at 10 seconds for now, but Omni Flash is already live inside the Gemini app, Google Flow, and YouTube Shorts, putting conversational video editing in front of every creator who already publishes there.
Every video model so far has asked you to write the perfect prompt the first time, then start over when the output drifts. Gemini Omni Flash flips that. You shoot once, then keep talking to the model.
If you are running an AI avatar account, this matters in two ways. The avatar mode lets you record your face and voice once, then generate clips from a chat box. And the conversational editing layer means the version that finally lands on Reels is the one you arrived at over five turns, not the first try.
This post walks through what shipped, how to use Omni Flash on your own account this week, and where it still does not beat HeyGen Avatar V or Hedra Omnia.
What's new in Gemini Omni Flash
Gemini Omni is a new multimodal model family from Google DeepMind. The first model, Omni Flash, started rolling out on May 19 to Gemini app users on the Plus, Pro, and Ultra tiers, to Google Flow, and to YouTube Shorts and the YouTube Create app at no cost. Developer API access is arriving in the coming weeks.
Three pieces matter for creators:
Conversational editing. You hand Omni an image, an audio clip, a video, or any mix of all three, then keep refining by chat. The model preserves character identity, physics, and scene context across turns. A violinist clip can be transported into a new environment, the violin can be made invisible, and the camera can move over the shoulder, all in one session.
Avatar mode. Onboarding asks you to record yourself speaking a sequence of numbers aloud. Omni then builds a personal avatar tied to your voice and likeness. General-purpose audio and speech editing on existing footage is being deliberately withheld at launch for safety reasons.
SynthID by default. Every clip carries an imperceptible SynthID watermark. You can verify it inside the Gemini app, Gemini in Chrome, and Google Search. This is the same provenance stack Google has been building for two years and matches the C2PA standard OpenAI rolled out earlier this year.
The Flash tier caps clips at 10 seconds at launch. Google has not disclosed per-clip pricing yet, or how Omni compares to Veo 3.1 or Seedance 2.0 on a benchmark suite. Expect both numbers once the API arrives.
Receipts: Introducing Gemini Omni (Google blog, May 19) and Google launches Gemini Omni Flash with avatar mode held back (The Next Web, May 20).
How to ship a Gemini Omni clip this week
1. Open Omni Flash. Find it inside the Gemini app on Plus, Pro, or Ultra; inside Google Flow on the same plans; or inside YouTube Shorts and the YouTube Create app at no cost during the rollout.
2. Drop in a reference. Upload a still of your avatar (or yourself), a short voice memo, and a five-second sample of a video style you want to copy. You can include all three in one prompt; Omni blends them into a single cohesive clip.
3. Write a layered prompt. Tell Omni the scene, the motion, and the audio behavior in one shot. Then refine in the chat box from there. The model treats every follow-up message as an edit on the current clip, not a fresh generation.
The following images were generated using Nano Banana 2:
Image prompt (Nano Banana 2): Editorial portrait of a mid-twenties AI creator walking through a neon-lit Tokyo arcade alley at night, oversized denim jacket, soft cinematic glow, 9:16 frame, identity-locked face, no text overlays.
Gemini Omni Flash video prompt: Apply the pose and walking motion from this reference video to the character in the attached image. Quickly style-shift from realistic cinema into anime, then back, twice, while the character keeps walking forward. Add background music that drops on each style change.
4. Refine across turns. Every follow-up is an edit, not a regeneration. Say things like "move the camera over her shoulder," then "change the jacket to ivory linen," then "add a slow-motion beat on the second style change." The character holds. The scene remembers what came before.
Image prompt (Nano Banana 2): Editorial mid-shot of an AI creator standing inside a bioluminescent forest at dusk, deep teal foliage, soft particle bokeh, 9:16, identity-locked face, no logos.
Gemini Omni Flash video prompt: Add a subtle harp glissando whenever the character touches a fern leaf. Make every fern leaf semi-translucent and pulse on each touch. Hold the room structure; only swap the lighting from dusk to dawn over 10 seconds.
5. Export with SynthID on. Every Omni clip ships with an invisible Google watermark. Leave it on. It is the cleanest provenance signal you can attach right now, and both Reels and TikTok have signalled they will privilege watermarked AI content under their 2026 labelling rules.
When NOT to use it
Omni Flash does not replace HeyGen Avatar V on a long-form talking head, and it does not match Hedra Omnia on full-body acted scenes longer than 30 seconds. The 10-second Flash cap is a real ceiling, since most reels you ship run 7 to 25 seconds, and Omni only nails the short end of that. Where it wins is the iteration loop. If your workflow is ship four reels a week and none of them lands on the first generation, Omni Flash inside YouTube Shorts is the only model right now that lets you stay inside the platform and keep editing by chat until the clip looks right. Use it for hooks, transitions, and styled b-roll. Use Avatar V or Hedra Omnia for the talking middle, then stitch.
Closing
If your account is built on an identity-locked avatar, your homework this week is three short Omni Flash clips: a hook, a transition, and a styled b-roll. Run them through your existing voice clone in Hedra Omnia for the spoken middle, then stitch the result for Reels and TikTok. The combined workflow is now the cheapest cinematic avatar pipeline on the internet, and if you do not have your own avatar yet, start there. The voice and likeness reference you record once is the unlock for every model in this list.
FAQs
Can I use Gemini Omni Flash for commercial content on Reels and TikTok?
Yes. The SynthID watermark is invisible and does not affect commercial rights. Both Reels and TikTok are moving toward privileging watermarked AI content under their 2026 labelling rules, so leave SynthID on.
Do I need a paid Google AI subscription?
For the Gemini app and Google Flow, yes: Google AI Plus, Pro, or Ultra. YouTube Shorts and the YouTube Create app get Omni Flash at no cost during the rollout.
How does Omni Flash compare to HeyGen Avatar V?
HeyGen Avatar V is built for long-form talking heads up to 30 minutes with phoneme-level lip sync in 175+ languages. Omni Flash is built for short, multi-input clips with conversational editing. They are complementary, not competitors. Most serious avatar accounts will end up running both.
Can I edit the audio or speech inside a clip Omni generated?
Not yet. Google is holding back general-purpose audio and speech editing for now while it tests the safety story. Avatar voice and music behavior are the only audio-side controls in the launch build.
When will the Omni API arrive for developers?
Google says the developer and enterprise API rollout is coming in the next few weeks. If history holds for Veo and Nano Banana 2, expect access through Google AI Studio first, with pricing tiers between Veo 3.1 and Seedance 2.0.