Gemini Generate Text

If you've been gluing LLM calls into ComfyUI with a requests node, this is the upgrade. Yogurt Gemini Generate Text is a real Gemini client inside the graph: it calls Google's Gemini API, keeps conversation history alive across executions, and even hands you the model's chain-of-thought as a separate output. No image generation here - the pack has a sibling node for that - this one is text in, text out, history for free.

What it is and why you'd reach for it

The Yogurt Nodes pack is a kitchen-sink collection, and this is its flagship LLM node. You wire it anywhere you'd otherwise put a hand-rolled HTTP call: rewriting a prompt for a second pass, generating captions to feed a captioning workflow, classifying an image description before it goes to a text-to-image node, or just asking questions mid-graph. Because it outputs a HISTORY type, it slots into the pack's other LLM nodes too - the same history object feeds its Image Understand siblings, so you can build a genuinely multi-turn pipeline without storing state yourself.

The inputs that matter

The list is long, but you mostly touch a handful:

• api_key - paste a Gemini key here, or leave it blank and let the node pick it up from config (see install).
• model_name - defaults to gemini-2.5-flash, which is the right call for most workflow use. Swap in pro when you need more reasoning.
• system_prompt and prompt - the system prompt sets the overall style; the prompt is your actual request. Both are multiline.
• temperature - higher = more random. Start at 1, dial down for deterministic labeling tasks.
• thinking_budget - this one's interesting: -1 means no cap on thinking, 0 disables thinking entirely. Pair it with the thought output below.
• seed - -1 for random, or pin a number to get reproducible-ish output. Handy when you're iterating on a prompt template.
• chat_template - the default is a {{system_instruction}} / {{prompt}} template with <‑system-> markers. If you use the history input, this is how each turn gets formatted, so it's the first place to look when a multi-turn conversation comes back garbled.

Also on the schema: top_p, top_k, max_output_tokens, retry_count, safety_level, disable_safety_settings, vertex (route through Vertex AI), base_url, proxy_url, timeout, and an extra JSON field for anything the client doesn't expose.

The outputs and what they wire into

• text - the model's reply (STRING). This is the one you save or feed forward.
• history - the updated conversation, ready to loop back into the same node's history input for the next turn.
• thought - the model's chain-of-thought as a separate string, when thinking is enabled. Great for debugging why a generation went sideways; don't ship it to a prompt that wants only the answer.

Install

The pack installs like any custom node. Fastest path is ComfyUI Manager - search "ComfyUI-YogurtNodes" and hit install. Or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/yogurt7771/ComfyUI-YogurtNodes.git
cd ComfyUI-YogurtNodes
pip install -r requirements.txt

Then restart ComfyUI. The LLM nodes need google-genai, which the requirements file pulls in.

For the key, the README gives you three options in priority order: paste one directly in the node's api_key field, drop it in custom_nodes/ComfyUI-YogurtNodes/yogurt_nodes/llm/api_key.json as {"gemini": "YOUR_API_KEY"}, or set GEMINI_API_KEY as an environment variable. Get a key from Google AI Studio. The file/env approach is better than hardcoding a key into every workflow - the key travels inside the workflow JSON otherwise, which is a bad habit to start.

Gotchas

Two things bite people. First, api_key is a required field, and if the file/env var isn't set, the node fails with an auth error - the README's priority order (node field > json file > env var) decides which key wins, so a blank field is fine only if the file or env var is configured. Second, remember this is a paid API: with max_output_tokens defaulting to 65,535 and thinking enabled, a long chain of workflow runs can quietly burn tokens. Set a sane token cap and a timeout so a hung request doesn't stall your whole batch.