Gemini Omni Video: Google's multimodal video model lands on MakeClipAI

A couple of weeks ago, I wrote about how choosing the right model depends on what stage your video is at. The models I covered — Kling, Seedance, Veo 3, Hailuo — all work roughly the same way: give them a prompt, maybe a reference image, and they generate a clip.

That's about to change.

Google's Gemini Omni Video just landed on MakeClipAI via the kie.ai marketplace, and it's the first model on the platform that genuinely thinks in multiple modalities at once. You're not just text-to-video anymore. You can feed it images, video clips, character IDs, and audio — all in the same request — and it weaves them into a coherent output.

I've been testing it for a few days. Here's what it actually changes about how I think about AI video prompts.

What makes "Omni" different

Most AI video models treat your prompt as a description. You write "a futuristic city at night with neon lights," and the model interprets that and generates something from scratch.

Gemini Omni doesn't work that way. It's trained to fuse multiple inputs simultaneously:

• Text prompt: The core description, same as any model
• Image URLs (up to 7): Reference images for character appearance, scene style, or storyboard frames
• Video clips (up to 1, ≤30s): A source video to remix, extend, or restyle
• Character IDs (up to 3): Character references from the gemini-omni-character API — keep a character consistent across generations
• Audio IDs (up to 3): Narration, dialogue, or sound design generated via gemini-omni-audio

The key difference: it can compose all of these together. An image reference for the character + a video clip for the background motion + an audio track for narration + a text prompt for the overall mood. That's not something the previous generation of models could do in a single pass.

The quota system is worth understanding

Because the model processes multiple inputs at once, the API uses a simple quota system. Think of it as having 7 slots:

• Each image consumes 1 slot
• Each video consumes 2 slots
• Each character ID consumes 1 slot

Formula: (Images × 1) + (Videos × 2) + (Character IDs × 1) ≤ 7

Practically this means:

• 7 images and nothing else
• 1 video + 3 character IDs + 2 images
• 5 images + 2 character IDs
• Or any other combination that fits within 7

This is actually pretty generous. Most use cases won't need more than 1-2 images anyway.

Where it shines

Character consistency is the biggest win. If you've used other AI video models, you know the pain of getting the "same" character to look the same across multiple shots. With Gemini Omni, you can pass a character reference via the character API, and it respects that reference across generations. This is huge for narrative work — multi-scene storytelling where the protagonist needs to be recognizably the same person.

Style transfer from video is another impressive use case. Feed it a 10-second clip of the visual style you want (specific lighting, camera movement, color grading), and it can generate new content that matches that style. The source video doesn't need to be high production value — even rough phone footage works as a reference.

Audio-guided generation is still early, but promising. You can generate dialogue or narration via the gemini-omni-audio endpoint and pass it in as an audio ID. The video output will sync reasonably well to the audio, which saves a lot of post-production lip-sync or voiceover alignment work.

Where it's not the best fit

Let me be honest about the tradeoffs.

If you're just doing simple text-to-video — "a cat playing piano" — Gemini Omni is overkill. You're paying for multimodal processing you don't use. Models like Seedance 1.5 or Kling 2.6 handle simple prompts faster and cheaper.

The same goes for rapid ad testing. If you're trying to churn through 20 hook variations in an afternoon, the quota system adds friction. You're better off iterating on Seedance or Kling and using Gemini Omni only for the final polished version.

Duration is also limited. The maximum output is 10 seconds. For longer scenes, you'll still want the Director mode with Seedance 1.5 or multi-scene Kling 3.0.

What this means for MakeClipAI users

Gemini Omni Video is available now in the model picker. You'll find it alongside Veo 3, Kling 3.0, Seedance, and Hailuo — same one-click generation workflow.

The pricing is comparable to premium models:

My recommendation: use it when you need multimodal inputs (character refs + audio + video). For standard text-to-video, stay on Seedance or Kling. Think of Gemini Omni as your "compose" model — the one you reach for when a single prompt and one reference image aren't enough.

Related reading

• Don't use the expensive model. Not yet.
• What I learned about picking AI video models after 200+ generations
• Stop rewriting AI video prompts from scratch every time
• How to Choose AI Video Models and Manage Credits