Dev.to AI ๐Ÿค– Ai ๐Ÿ‘ 0 ๐Ÿ“– 5 min read

Gemini Omni 1.2 Flash for Builders: What to Test Before You Rely on 15-Second AI Video

Google released Gemini Omni 1.2 Flash on November 5, 2026, with native video generation and editing up to 15 seconds, reference-material following and multi-turn edits. If you are thinking about putting AI video generati

Gemini Omni 1.2 Flash for Builders: What to Test Before You Rely on 15-Second AI Video

Google released Gemini Omni 1.2 Flash on November 5, 2026, with native video generation and editing up to 15 seconds, reference-material following and multi-turn edits. If you are thinking about putting AI video generation or editing into a product or a production pipeline, the useful question is which of those claims survive a test.

Two outside testers have published results, and one run of my own adds a data point from Veo 3.1:

  • Treza (November 10): three reference images kept a robot's main design consistent across four shots. It floated once and briefly grew a mouth.
  • DataCamp (November 6): a color edit kept the action and event timing, and kept the original clip's glitch.
  • My Veo 3.1 run (November 16): the job completed through the API, and a second marble appeared at second 7.

Who this is for: teams planning character storyboards or recolor-and-restyle edits on clips they already like. Who it is not for: anyone who needs physically exact output. I could not find Omni in Crazyrouter's catalog, so Omni evidence here is third-party. The Veo run is a separate task, so none of this is a head-to-head.

Three claims, three acceptance checks

Here is each claim mapped to its evidence and to an assertion you could put in your own acceptance test.

Claim 1: reference images keep a character consistent

  • Evidence: Treza, three references, four 8-second shots. Main design held; floating and a brief mouth remained.
  • Acceptance check: compare every shot against a character sheet, then watch the moving frames separately for pose and ground contact.

Claim 2: you can edit without redoing the clip

  • Evidence: DataCamp, 15-second clips. Red-to-blue kept position, appearance and event timing; the vanishing-marble bug stayed.
  • Acceptance check: compare event timestamps before and after each edit, flag inherited defects, and measure audio offset per pass.

Claim 3: longer clips are usable

  • Evidence: Google's announcement lists fast motion, rapid plot changes and complex physics as areas still improving; DataCamp's 15-second marble run had an object vanish and return.
  • Acceptance check: count the main objects at sampled frames through to the last second.

What changed in Gemini Omni 1.2 Flash

The announced workflow is draft, refine, then upscale.

  • Length: 3 to 15 seconds in 1-second steps; the previous limit was 10 seconds.
  • References: character, product or style material; which inputs work depends on the access route.
  • Edits: repeated requests on the same clip for background, lighting, weather.
  • Resolution: 360p or 720p drafts, upscaled to 1080p or 4K.
  • Audio: supported; none of the cases below involve dialogue.

DataCamp notes that the public Gemini API does not accept uploaded audio references and that some Google Cloud features need approval. Check the fields and durations of your actual route before designing around the full list.

Claim 1 in detail: reference images and four shots

Alex Daro's Treza write-up used a lantern robot.

  • Inputs: front full-body, face close-up and side full-body images.
  • Outputs: four 8-second videos, 720p, 9:16, sound on.
  • Shots: wide push-in, close-up, side tracking, low angle.
  • Control: the same prompts with the references and matching prompt lines removed.

In the control set, the lantern body, copper head and teal scarf appeared every time, yet the face, limbs and build changed shot by shot. Add the references and the core design stopped drifting.

Treza's Omni reference-image test: the lantern robot in the first shot

Image: Alex Daro / Treza, November 10, 2026. First shot of the with-reference set; original author's Omni output.

Two problems remained. After letting go of the lantern cord the robot hovered, and the author sidestepped it by rewriting the prompt to have both feet planted on a crate. Separately, a design with no mouth showed one for part of a shot. One stylized character, tested by a company that also sells model access.

Claim 2 in detail: color and style edits

Khalid Abdelaty's DataCamp test ran three passes on 15-second, 720p, 16:9 clips.

  1. Original: a red marble rolls along a track, through a copper ring, knocks over dominoes and rings a bell. The marble disappears and reappears.
  2. Red to blue: position, main appearance and event timing mostly held, and so did the disappearing marble.
  3. Stop-motion style: cardboard texture on the track and dominoes, parts of the background still realistic, motion still smooth.

Reported numbers: under a minute for the first generation, about 2 minutes per edit, audio about 45 ms late after the first edit and about 85 ms late after the second. Treat each edit as a new job in your timing budget and re-measure audio each time.

Claim 3 in practice: what my Veo 3.1 run showed

I wanted a first-hand look at object continuity. On November 16 I checked the Veo 3.1 page on Crazyrouter and sent one text-to-video request to https://api.crazyrouter.com/v1/video/generations with model ID aigc-video-gv-3.1.

  • 8 seconds, 720p, 16:9, sound on
  • One submission, no retries
  • About 67.35 seconds from submission to the poll that reported completion
  • File: 1280ร—720, 24 fps, AAC stereo at 48 kHz

Prompt, paraphrased: a red glass marble in a warm clockmaker's workshop rolls down a wooden ramp, strikes the copper bell on the bench and comes to rest next to it. Single unbroken shot, both objects in view at the moment of contact, nothing duplicated or disappearing, a rolling sound and a bell tone on impact, and no music, speech or captions.

I dropped DataCamp's ring and dominoes so the marble was the only thing to track.

Veo 3.1 frame at second 0: red marble, wooden ramp and copper bell

Second 0 of my clip. The next two frames come from the same file.

Veo 3.1 frame at second 4: one marble left of the bell

Second 4: one marble, left of the bell.

Veo 3.1 frame at second 7: two red marbles

Second 7: two marbles, after a prompt asking for one and ruling out duplicates.

The submit, wait and download path worked, and the opening frame looked right. A check that sampled only early frames would have passed this clip. The file has an audio track; I did not measure audio-video sync, so I am not rating the sound. To try a similar job, see the Veo 3.1 model and call details.

Where each model fits

  • One character across several shots: try Gemini Omni 1.2 Flash first, with a consistent set of views, and re-check the character in motion.
  • Keep the action, change color or style: try Omni first; change one thing per edit and save every version.
  • A standalone draft shot, already working through Crazyrouter: Veo 3.1. The whole job ran end to end; inspect subject count and the last seconds. Google's other models are listed in Crazyrouter's Google model matrix.
  • Exact physics or collision teaching: film it or use controlled animation. DataCamp's original and my Veo clip both broke object continuity.

Pick by the job and by what you can reach today. If you have run Omni on a physics shot with references attached, I'd like to hear whether the object count held.

Sources

  1. Google, November 5, 2026: "Gemini Omni 1.2 Flash: Fast and cost-effective with longer clips."
  2. Treza Labs, Alex Daro, November 10, 2026: "One Character, Four Angles: Gemini Omni 1.2 Flash With Reference Images." Treza also sells model access; its results are credited as published.
  3. DataCamp, Khalid Abdelaty, November 6, 2026: "Gemini Omni 1.2 Flash: New 15-Second Clips and Lower Pricing."

Third-party results and release details checked November 16, 2026. My Veo 3.1 run took place the same day.

๐Ÿ“ฐ Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes โ€” full credit and traffic to the original publisher.