Gemini Omni Video

Most video nodes take a prompt and maybe a start frame. This one takes a prompt, an image batch, trimmed source clips, registered characters and voice IDs, then lets you ask for 4K. It's the assembly point for everything else in the Gemini Omni group, and the budget rules are the part you need to read before you queue anything.

What it is

A text/image/character-to-video node on KIE's Gemini Omni API. The reason it exists rather than a generic API node is the pack's whole design decision - one ComfyUI node per KIE model/operation, with the provider's documented request fields turned into real widgets and the URL plumbing replaced by native IMAGE, VIDEO and AUDIO sockets. Connected media gets uploaded to KIE automatically.

That design is why people put up with a marketplace in the first place. The community's argument for KIE is price against Comfy's official API nodes - public comparisons have put some Google models roughly 40–60% cheaper through KIE's route - and the trade is that you're calling a reseller with its own model churn instead of a first-party integration. Both things are true; pick accordingly.

Inputs

prompt (required) - the shot, in prose. The default is a night-city push-in, which tells you the register they expect: camera language, action, subject.

duration - a string, 4 by default. Documented values are 4, 6, 8 and 10 - and the tooltip is explicit that when you supply video input, the model decides the output length and this setting is ignored.

image_urls - a native IMAGE socket. Your reference images.

resolution - 720p, 1080p or 4k. Pick deliberately; this is the biggest cost lever on the node.

aspect_ratio - 16:9 or 9:16. That's it.

seed - INT, 0 by default. KIE notes results can still vary at a fixed seed, so treat it as a nudge, not a reproducibility guarantee.

audio_ids - a JSON array string of voice IDs from the Omni audio endpoint. Max 3 items. Good for narration or dialogue.

character_ids - a JSON array string of IDs from the Omni character endpoint. This is where the budget bites: each character consumes 1 image slot, the base limit is 7 image slots, and if you also pass a video_list that eats 2 slots, leaving fewer for characters.

video_list - a JSON array of clips to work from. Each entry takes a URL plus a trim range (start / ends). Documented limits: 100MB per file, 30 seconds max duration, and ends has to be greater than start.

timeout_seconds - 1200 by default, 30–7200.

callback_url - leave empty for normal ComfyUI use.

Outputs

video, url, all_urls_json, task_id, raw_json, credits_consumed, credits_left. The video socket is a real ComfyUI VIDEO object, so it goes into a Save Video, a preview, or another node that accepts video. url is the primary result, all_urls_json keeps everything the API reported (useful when a job returns extra artifacts), and the credit floats plus the on-node receipt tell you what the render cost.

Install

ComfyUI Manager, search the pack title, or:

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next
pip install -r ComfyUI-KIE-Nodes-Next/requirements.txt

Restart. Dependencies: requests, Pillow, numpy, PyYAML. Nothing renders locally - this node is a client, and the GPU work happens on KIE's side at per-call prices.

Key: Settings → KIE.ai Nodes Next → Connection → Paste / replace KIE.ai API key; validated before saving, masked fingerprint afterwards. Headless: KIE_API_KEY wins over the saved key.

Gotchas

Media slots are the trap. Characters and source clips compete for the same 7-slot image budget. Two characters plus a video_list is 4 slots before you've added an image reference.

JSON fields are strings. audio_ids, character_ids and video_list are text fields containing JSON arrays. Malformed JSON fails locally rather than at KIE - which is the good outcome, but it still means the node errors before spending anything.

Duration is a red herring with video input. It's ignored by design when a source clip is present. Don't tune it and wonder why the length didn't change.

4K takes time. Raise timeout_seconds if you're queuing a long 4K job rather than assuming the run hung.

The card list is provider-driven. Nodes are generated from KIE's live API documentation at startup with a daily re-check, so field-level changes arrive with a restart, and the occasional model disappearing is upstream, not your install.