✨ Gemini Configurator

The Gemini Configurator doesn't do anything on its own, but you can't run this pack without it. It's the node that holds your API key, picks the model, and bundles all of that into a single gemini_model connection that the 🚀 Gemini API Runner consumes. Think of it as the key safe plus the model switch for everything Gemini in the workflow.

You wire it once at the top of your graph: gemini_model output → runner. That's the whole job.

The inputs that matter

Four inputs, and honestly only two of them will occupy you most days:

• api_key - your Google AI Studio key. It's a plain STRING in the backend, but the pack's frontend JS re-draws the widget so it renders as password dots instead of visible characters. That's a genuine nicety if you screen-record or share your canvas: your key stays masked. If you leave it empty, the configurator raises a ValueError the moment you run - it's the one hard stop in the pack.
• model_name - the big decision, and the list tells you what Google's API serves these days. Default is gemini-2.5-flash, a fast cheap workhorse. Below it sit gemini-3-flash-preview and gemini-3.1-flash-lite-preview (newer, better thinking), the tiny gemini-2.5-flash-lite, three Gemma open-weights models (gemma-4 and gemma-3) served through Google's API, and the two native-audio live models that power the pack's voice feature. Which one you pick changes what the runner can do - see below.
• voice_name - the five Gen-Audio personas: Puck, Charon, Kore, Fenrir, Aoede (default). These are Google's dramatic-reading voices, tuned for character rather than the robotic drone of typical TTS. They only matter if your chosen model actually generates audio, so don't sweat this for text-only models.
• thinking_level - default, minimal, low, medium, high. Straight from the source, this only takes effect on Gemini 3+ and Gemma 4 models; for others the runner ignores it. Level up if you want better reasoning and don't mind the latency.

The model-selection trap

The pack's headline is voice, but the default gemini-2.5-flash is text-only. If you want the runner to return real audio, you must pick a gemini-3 model (which gets TTS through Google's speech config) or one of the two native-audio/live models. And note the reverse: those native-audio models don't accept images. So there's a real fork here - vision analysis wants the standard text models, voice wants the audio models. The configurator is where you make that call.

Outputs and reload

One output: gemini_model (GEMINI_MODEL), a dict carrying the client, model name, key, voice, and thinking level into the runner. There's also a "🔄 Reload Model" button added by the frontend - the pack's recommended way to apply a key change without restarting ComfyUI. It just refreshes the config and tells you to re-queue.

Install

Standard for this pack: ComfyUI Manager → search "Universal LLM Suite" → Install, or clone it into custom_nodes and restart:

cd ComfyUI/custom_nodes/
git clone https://github.com/alice-ai-wonder/ComfyUI-Universal-LLM-Suite

On first boot it auto-pip-installs google-genai, openai, and librosa. No local model to download - this is a pure API pack, so the heavy lifting (and the bill) lives at Google's end. Grab a key from Google AI Studio, paste it in, and you're set.

Common issues

• "API key must not be empty" - you skipped the field; the node refuses to configure.
• Runner shows "❌ Gemini API Error" - usually a bad/expired key or a quota problem; the runner doesn't crash, it returns the error as text, which is actually the easiest way to debug it.
• Audio comes back silent - you left the model on gemini-2.5-flash. Switch to a gemini-3 or native-audio model.

It's a small, young pack (one commit, zero community footprint as of mid-2026), and like any custom node it runs arbitrary Python on load - the usual caveat applies: glance at the source once, don't just trust the convenience.