Short vertical ads live or die in the first two seconds, and most AI video models now produce clips that are long enough for a hook but too short for a whole story. Gemini Omni Flash 1.1 is a good example: it renders 4, 6, 8 or 10-second clips, in 16:9 or 9:16, at 720p, 1080p or 4K. That shape suits paid social well if you plan for it instead of fighting it. Here is a simple way to plan for it.
1. Decide the clip length before you write the prompt
Pick the length from the job the clip has to do. A 4-second clip is a hook or a product reveal. Six to eight seconds fits one action plus a beat of reaction. Ten seconds is enough for a simple before-and-after. Writing the prompt for a specific length stops the model from cramming three ideas into one shot.
2. One subject, one action, one camera move
Short clips look best when the prompt names a single subject, a single action and a single camera instruction ("slow push-in", "locked-off overhead"). If you need a second action, make it a second clip and cut them together.
3. Use first and last frames for product shots
For image-to-video you can upload a first frame and, optionally, a last frame. For product ads this is the easiest way to keep packaging, logo and colour consistent: shoot or render the product still, use it as the first frame, and describe only the motion.
4. Plan the audio separately
Gemini Omni Flash 1.1 does not generate audio, so treat sound as an edit step: pick the music bed and voiceover first, then time the clips to it. The model can also render animated on-screen text, which helps for sound-off viewing in feeds.
5. Render at 1080p first, 4K only for the keeper
Iterate at 720p or 1080p, then re-render the version you will actually run at 4K if the placement needs it. It keeps the cost of exploration down.
6. Test one variable at a time
When a clip works, change only the hook line or only the first frame for the next variant, so you know what moved the result.
Disclosure: I work on marketing at Veo3 AI, an independent AI video platform that is not affiliated with Google. If you want to see the clip lengths, formats and prompt examples in one place, the Gemini Omni Flash 1.1 page on Veo3 AI lists them, along with a note that the original Gemini Omni model was retired from the generator on September 1, 2026.
