Gemini Omni Flash

If you've only ever seen text-to-video, Gemini Omni Flash is the model that makes you re-read the word "omni." It's Google's multimodal video model, and the node's input list is the tell: images, frames, an existing video, a videoId to build on - it swallows all of them and emits a video. This isn't the "type a sentence, get a clip" tier. It's the "turn these reference photos into a coherent moving scene, or edit footage you already have" tier.

Because it's a Runware pack node, none of this happens on your machine. The node serializes everything you feed it and ships it to Google's model over Runware's cloud API, then hands the finished clip back as a native VIDEO output. No GPU, no weights, no FFmpeg dance on your end - that's the whole value proposition of this pack, and it's why people without a 24GB card reach for these nodes at all.

What you actually set

The only required input is positivePrompt - and the tooltip is worth reading, because this model accepts aspect ratio and duration in the prompt itself ("a 16:9, 8-second shot of…") as an alternative to the widgets. The interesting inputs are the media ones:

• referenceImages - stills that define subjects and style.
• frameImages - the keyframe stack; the model animates between your frames rather than inventing motion from nothing.
• video - a URL, path, or UUID of footage to edit. This is the one that turns the node from a generator into an editor.
• videoId - pass back the ID of a previous generation to iterate on it in a follow-up turn. Underrated for tightening a shot: generate, look, tweak the prompt, hand the ID back, get a refinement of the same idea instead of a lottery ticket.

Then the usuals: width/height (default 1024²), duration behind a toggle (3–10 seconds - flip the toggle on, then set duration_value), temperature and topP if you want more or less randomness, and numberResults up to 4. Output is a single video.

The gotcha shared by every Runware node

Notice the duration pattern: it's a boolean toggle, and the actual value lives in duration_value. This pack gates every model parameter that has a validation rule behind an on/off switch, so "Off uses the model's default." A beginner's first run produces a default-duration video because they typed a number into the disabled duration_value field and never flipped the toggle. Flip the switch, then set the value.

Beyond that: you need a Runware API key (dashboard → Settings → Runware API key, or RUNWARE_API_KEY), and every run costs money - the node prints the price on its title bar, so watch what you're spending when you set numberResults to 4. Video is the expensive end of this pack, and iterating is where the bill creeps.

Install

Same as every node in the pack - ComfyUI Manager, search "Runware," install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

No model downloads, no CUDA version roulette. The one thing the pack does need is the key. Wire the video output into a Video Preview or Save Video node and you're running a model that would otherwise be completely out of reach locally.