Google Gemini Omni (Video)

This is the odd one out in the partner/video family, and that's exactly why it's worth a look. Most of these nodes give you a box full of widgets - resolution, duration, aspect ratio. Gemini Omni gives you a text field and says "figure it out." It's a hosted API node built into ComfyUI core that calls Google's Gemini Omni Flash model, and every setting that other nodes expose as a dropdown, this one expects you to describe in the prompt: "a 6-second clip in 9:16, drone pulls back to reveal a lighthouse at dusk." Length must be 3–10 seconds, aspect ratio 16:9 or 9:16, output is 720p at 24fps with audio. All of that is prompt-shaped, not widget-shaped.

The reason it's called Omni is the reason to care: it's one model that generates video and synchronized audio together, like Veo 3 - dialogue, ambient sound, the works. No separate audio step, no post-sync. And because it's a multimodal model, the way you steer it is broader than "image in, video out." Click into the model dynamic combo and you get sub-inputs: prompt, up to 14 reference images (autogrow), up to 3 reference videos (each max 10 seconds), plus temperature and top_p for the sampling behavior.

That's the real power move. Feed it a reference video of a scene or subject and a prompt asking for a change, and you're doing video editing, not just generation - same family as the image-editing Gemini nodes, but in motion. Reference media is optional; a bare prompt works fine and is the simplest way to start.

Outputs are the unusual part: you get VIDEO and a STRING. That second output is the model's text response accompanying the clip, which you can read for context or wire into a text display. If you're scripting an automated pipeline, it's a genuinely handy detail - you get the video and the "here's what I did" note in one shot.

Gotchas, and they're real. First, non-determinism: seed (default 42) only controls whether the node re-runs; the output will differ anyway, so don't chase a specific result with seeds. Second, the content rules - this is Google's model, so expect it to refuse some prompts outright, and when it does, the failure is an error naming the blocked content rather than a graceful skip. Third, the price: it's billed per second of output, and the badge the node shows is a per-second rate that adds up fast on a 10-second clip. The old community habit - draft cheap, then spend - applies harder here.

It landed in core at the end of June 2026 and was quietly reworked in mid-July to route through Gemini's Interactions API, which is why reference videos work as well as they do. If you want Google-grade audio-synced video and you're comfortable writing a prompt that does the work of ten sliders, this is the one. If you'd rather have explicit controls, the other nodes in the family are waiting with their combos.