Omni 1.1's 3-second video reference

Omni 1.1 Flash’s video reference is up to three seconds of footage, not a folder of twelve files. Google’s 27 August 2026 Keyword post says you can reference up to three seconds of video when crafting a scene, to keep visual context and character consistency from video — not only from stills. The demo replaces dancers in uploaded clips with provided characters, then restages the dances in a new space. That is a short motion citation. It is not Seedance 2.0’s multi-reference stack (text, image, video, audio; up to twelve references in ByteDance’s 2.0 writeups). Versely does not publish /models/gemini-omni-1-1-flash. The Omni rows we publish are Gemini Omni Video and Gemini Omni Flash Edit.

This page is only the three-second citation versus a multi-file brief. Motion control is a different family.

Three seconds is a clip, not a bible

A video reference is pixels that already move. The model can copy timing, weight, and a face that exists across frames instead of inventing them from a still and a paragraph. Google capped that citation at three seconds on 1.1. Three seconds is a gesture, a step, a turn of the head, a product flip. It is not a twelve-second dance. It is not a trailer. If your reference is longer, you are asking 1.1 to ignore most of it or to fail the citation.

Google’s dancer example is the right shape: short source performances, character stills, a location still, one continuous output. The video files supply the choreography. The stills supply who. The prompt names the mapping (“the dog does the classical dance from dance3.mp4”). That mapping is the brief. Without it, three seconds of video is just another noisy conditioner.

What a 3s ref is good for:

• A specific gesture you cannot spell in prose (the exact shrug, the exact pour angle).
• Identity that is easier to hold from moving frames than from one hero still.
• Restaging a short bit of motion onto a new character or a new room.

What a 3s ref is not:

• Kling motion control. Kling Video V3 Pro Motion Control transfers a driving clip onto a reference image at 1080p, from 68 credits (14 a second), audio off. That row’s job is puppeteering. Omni 1.1’s 3s ref is multimodal generate with a short video among the inputs.
• A Seedance 2.0 reference pile. Seedance 2.0 on Versely is Seedance 2.0: 4–15s, 1080p, from 44 credits, native audio, text-to-video and reference-to-video. ByteDance’s 2.0 stack is documented as up to twelve references across stills, video, and audio. That is a brief made of files. Omni’s 3s is one short clip’s worth of motion context.
• First-last interpolation. Two stills as endpoints are a different 1.1 control. A 3s video is not a last frame.

Trim the reference before you attach it. If the useful motion is 0:01–0:04 in a thirty-second screen recording, cut to those three seconds. Feeding a long take and hoping 1.1 “finds the bit” is how the citation becomes a blur.

Omni Video’s quota is not the same number

On Versely, Gemini Omni Video already accepts a prompt plus optional reference images, source video clips, character IDs, and audio IDs. The published quota is images + videos*2 + character_ids <= 7. Durations are 4s, 6s, 8s, 10s. Output lists 720p / 1080p / 4K. Credits start at 29. That card is how you run Omni-family generate here. It is not Google’s 27 August sentence “reference up to three seconds of video.”

Do not collapse the two:

• Omni Video quota is a slot budget. Videos cost two slots each. A couple of clips plus stills plus character IDs can fill 7 without any of those clips being capped at three seconds in the catalog text.
• Omni 1.1 video ref is Google’s 1.1 control: up to three seconds of video as the cited motion/identity context, on gemini-omni-1.1-flash, in AI Studio / Agent Platform / Flow.

If you are on Versely and you needed Omni-shaped generate with refs, use Omni Video and respect the quota. If you are on Google’s 1.1 surface and you needed the three-second citation as Google described it on 27 August, use that control. Do not invent a third slug that merges them.

Gemini Omni Flash Edit does not take a 3s reference to start a scene. It takes an existing video and an instruction, from 50 credits, audio off on the card. Edit is retouch. Reference is casting-plus-motion. If the take already exists and the jacket is wrong, Edit. If the take does not exist and you have three seconds of the move you want, 1.1 video ref or Omni Video, depending on the door.

Seedance’s stack is a different brief

Creators keep pasting “12 references” onto every multimodal model. That number is Seedance 2.0’s, from ByteDance’s 2.0 writeups (text, image, video, audio; up to twelve references). On Versely, run that family on Seedance 2.0 or the fast reference cousin Seedance 2.0 Fast Reference to Video (4–15s, 720p, from 55 credits, was 78, native audio, discounted). Read the card. Do not assume twelve slots on Omni.

Seedance 2.5 on Versely is text-to-video: 4–30s, 720p, from 34 credits (was 67), native audio. The published category list is text-to-video, not reference-to-video. A 30-second single pass is a continuous action from a prompt. It is not a 12-file brief and not a 3s Omni citation. If you needed 2.5’s length, write one verb. If you needed a stack of files, you are on 2.0 reference, Veo R2V, or Omni Video’s quota — pick one.

Veo’s stack is smaller and stricter: VEO 3.1 Reference to Video takes up to three stills, 8s only, from 128 credits, native audio. No driving video on that card. Three photos, new cinematic eight seconds. If your “reference” is footage, Veo R2V is the wrong row.

Kling motion control is the driving-clip specialist: Pro from 68 credits (14 a second), 1080p; Standard from 51 credits (11 a second), 720p; both audio off; both require the still and the motion. When the joke is the exact choreography, Kling. When the citation is three seconds of look-and-move inside an Omni generate, Omni 1.1’s video ref.

Wrong-door table, short:

You have	You want	Row / surface
≤3s of motion + stills	Omni-looking restage	Omni 1.1 video ref (Google) or Omni Video quota (Versely)
Up to ~12 mixed refs	Seedance 2.0 multimodal	Seedance 2.0 / 2.0 Fast R2V
Three stills, new 8s cinematic	Veo recast	Veo 3.1 R2V
Driving clip + character still	Copied choreography	Kling V3 motion control
Existing Omni take, one fix	Retouch	Omni Flash Edit

How to cite three seconds without lying to the model

Cut the ref to the action. Three seconds of a clean gesture beats twelve seconds of a messy take that contains the gesture.

Light the stills like the ref, or accept a restage. Google’s dancer demo restages into a new space on purpose. If you needed the original room to survive, you wanted image-to-video or an edit, not a citation into a new scene.

Name the mapping in the prompt. “Use clip A for the pour, still B for the bottle, still C for the hands. One continuous shot.” Unmapped refs fight.

Do not also ask for a forty-second extend in the same breath. 1.1 extend uses up to ten seconds of prior generated context to write the next ten seconds, to a forty-second cap. That is a sequel control. Video ref is an input control. First lock a 4–10 second generate that honored the 3s citation. Extend the keeper later if the story needs length.

Draft at 360p on 1.1 if you are still hunting (Google: up to 60% faster and one-third the cost of 720p — Google’s claim, 27 August 2026). A 3s ref does not make a 4K draft cheap. Citation first, resolution second.

Caption after. Add captions to video. The ref will not typeset.

The AI video generator is the Versely door for Omni Video. For 1.1’s three-second citation as Google shipped it, use Google’s surface and bring the file back.

If the job is Omni generate on this catalog, sign up and open Omni Video — and keep Seedance’s twelve-file stack on Seedance.

FAQ

Is Omni 1.1’s 3-second video ref the same as Seedance’s 12 references?

No. Google’s 27 August 2026 post caps Omni 1.1 video reference at three seconds of footage. Seedance 2.0’s writeups describe up to twelve mixed references. Different families, different briefs. Run Seedance on Seedance 2.0.

Does Versely’s Omni Video row cap refs at three seconds?

The published Omni Video card uses a slot quota (images + videos*2 + character_ids <= 7), 4–10s, from 29 credits. It does not reprint Google’s “up to three seconds” sentence. Do not invent a merged spec. Read the card you are confirming.

Can I use a 15-second dance as the Omni 1.1 video ref?

Not as specified. Google’s cap is three seconds. Trim to the phrase you need, or use Kling Video V3 Pro Motion Control when the whole performance is the content.