Blog / 5 min read · 2026-05-24

Gemini Omni Avatar feature — hard rules, recording setup, and what actually works

Everything verified about Gemini Omni's Avatar feature: age + region gates, identity verification, recording requirements, SynthID watermark, and the @username summon syntax.

The Avatar feature is Gemini Omni’s most novel capability — summon your own face and voice into a generated video via the @username syntax. It’s also the most regulated feature Google has shipped in Omni, with hard gates on age, region, and identity verification.

This post collects every verified rule from Google’s help page and the public hands-on reviews, plus the recording requirements that determine whether your avatar works or not.

The hard rules

Set by Google’s Avatar help page. All of these are non-negotiable at launch (May 2026):

• Age — 18+ required. Google verifies via the account birthdate, not self-declaration.
• Geography — Not available in EEA, Switzerland, or UK at launch. VPNs don’t bypass; Google checks account region and IP.
• Language — English only. Voice cloning and prompt-to-avatar mapping work only in English at launch.
• Subscription tier — Avatar is gated behind paid Google AI plans (Plus / Pro / Ultra). Not available on the free YouTube Shorts surface.
• Identity verification — A live face + voice recording is required during setup. This anchors your avatar to your real identity in Google’s system.
• SynthID watermark — Every avatar-generated video carries Google’s SynthID identifier embedded in pixels. Not optional, cannot be removed, persists through compression and format conversion.

The reference recording — what makes or breaks your avatar

This is where most setups fail. Per Chrome Unboxed’s hands-on, the recording determines facial accuracy, lip-sync quality, and emotional range. Get this right once and your avatar works for months.

Camera + framing requirements

• Clear view of eyes, nose, and mouth. No glasses with heavy reflections. No partial occlusion.
• No sunglasses, masks, or hats covering the face. Even partial.
• No other faces in the background. Single subject only.
• Good even lighting on the face. Front-lit or three-quarters-lit beats side-lit. Avoid harsh shadows across one half of the face.
• Face fills roughly 40-60% of the frame. Too close = distortion at edges. Too far = not enough detail.

Recording duration + content

• ~30 seconds of varied expression. Smile, neutral, slight head turn, blink naturally.
• Speak ~5-10 short sentences for the voice clone. Vary intonation. Don’t read robotically.
• Look at the camera. Eye contact is encoded; recording with eyes averted produces avatars that won’t make eye contact in outputs.

Common failure modes

• Recording while wearing the glasses you usually wear. Your avatar will always wear glasses in outputs, even if your prompt says otherwise.
• Recording with a strong haircut/style. Same — that style gets locked in.
• Background clutter or another person walking past. Google’s preprocessor sometimes confuses the subject.

If any of these happen, re-record from scratch. The avatar isn’t easily editable after creation.

The @username summon syntax

Once your avatar is approved, use it in any Omni prompt with @your_username:

@your_username sits at a wooden desk in a warmly lit home office,
talking to the camera. 10-second 16:9, one continuous shot,
slight handheld, 35mm film grain.

Notes:

• @your_username must match your Google account username exactly. Case-sensitive.
• Only one @username per prompt. Multi-avatar scenes don’t work at launch.
• You can combine @username with @character_name reusable characters. Useful for “you + your dog” or “you + a fictional companion” scenes.
• All standard prompt rules still apply. Camera vocab, duration locks, “keep X identical” all work with avatar prompts.

What avatars are good at

From multiple hands-on reviews:

• Talking-head shots — strongest use case. Lip-sync is “shockingly accurate” (Chrome Unboxed) for English speech in well-lit indoor scenes.
• Reaction expressions — surprise, concern, smile, laugh. Micro-expressions render well.
• Subtle camera movement during speech — slight handheld, slow push-in, soft pan. Big moves drop accuracy.
• Simple background swaps — same talking-head, different office / coffee shop / outdoor scene.

What avatars break on

• Side profiles — accuracy drops sharply past 45° head turn from frontal.
• Action sequences — running, dancing, complex hand articulation degrade fast.
• Hands holding objects near the face — phone-to-ear shots tend to look uncanny.
• Multiple people in frame — second face often becomes blurry or warped.
• Non-English voice prompts — voice clone reverts to a default English voice or breaks lip-sync entirely.

Compliance + ethical guardrails

Google has baked in several safety patterns worth knowing:

• You can only summon your own avatar. Attempts to clone another person’s face from a photo are rejected at the verification step.
• SynthID is mandatory. Any avatar output can be identified as AI-generated through Google’s verification tools.
• Reuse limits. Google has stated avatar usage is monitored for misuse; commercial deployment beyond personal content may require additional review.

For most independent creators, this means: use it for your own content, attribute clearly, don’t try to impersonate anyone.

Quick setup checklist

Before recording your reference:

• Plain background, single subject (you)
• Front or three-quarters lighting, no harsh shadows
• No glasses (unless you always wear them), no hats, no masks
• Face at 40-60% of frame
• 30 seconds of varied expression
• 5-10 spoken sentences with varied intonation
• Look at the camera

Before submitting your first avatar prompt:

• Account confirmed 18+
• Not in EEA / Switzerland / UK
• Paid Google AI tier active
• Reference recording approved
• Prompt uses English only

Related

• Avatar prompts in the library — production-ready avatar templates
• “Keep X identical” lock — applies to avatar multi-turn editing
• Full field guide — broader Omni context
• Glossary entries — formal definitions for @username, SynthID, Avatar