Vertex AI Generate Text

You've got a sampler that needs a good prompt, and you've got an idea that needs a writer. YogurtVertexAIGenerateText is the pack's general-purpose Gemini text node: feed it a system prompt and a prompt, and it returns generated text you can wire straight into a CLIP text encode. What makes it different from a dozen similar nodes is the auth flexibility - it can talk to Gemini through a plain AI Studio API key or through Google Cloud's Vertex AI, and you flip between them with a single toggle.

How it works

The model defaults to gemini-2.5-flash, a fast and cheap pick that's right for prompt work. The node goes through the google-genai SDK (the google-genai dependency in the pack's requirements) and outputs:

• text - the generated response, a plain STRING
• history - the conversation so far, for chaining turns
• thought - the reasoning trace when the model is in thinking mode

Optional history input lets you feed a conversation back in for multi-turn work (the pack's YogurtHistoryBuilder exists for exactly that), and extra accepts a JSON string of extra request parameters.

The toggle that matters: vertex

This is the node's whole personality. Set vertex to false and it authenticates with api_key (a Gemini API key from Google AI Studio, optionally routed through base_url). Set it to true and it switches to the Vertex path: credentials (a GCP service-account JSON), project_id, and location instead. So this one node covers both routes that the pack's other Gemini nodes split across. Start with vertex off and an AI Studio key - it's dramatically less setup, and for prompt generation you don't need the GCP plumbing.

Inputs worth your attention

• system_prompt (multiline) - how the model behaves. "You write high-quality Stable Diffusion prompts" is the kind of thing this is for.
• prompt (multiline) - the actual ask.
• temperature (default 1) - higher, more random; for prompt generation, values under 1 tend to give steadier results.
• max_output_tokens (default 65535) - the cap on the reply.
• thinking_budget (default 0) - 0 disables thinking; -1 means no limit. For creative prompt work, a small thinking budget can noticeably improve output without much latency cost.
• seed (default -1) - set a fixed seed for reproducible text.

Where you'll use it

The obvious one: prompt augmentation. Generate a batch of prompt variations, feed them through YogurtStringSplit into a loop, and let the sampler churn through text the model wrote for you. Or the slightly more advanced pattern - an LLM "quality pass" that rewrites your raw idea into a structured prompt before it ever touches the CLIP encoder. Since the output is a plain STRING, it plugs into everything.

Installing it

Part of the Yogurt Nodes pack (YogurtNodes/LLM), installed once:

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

Or ComfyUI Manager, search "ComfyUI-YogurtNodes." Restart after. For the API-key path, set GEMINI_API_KEY or drop an api_key.json into custom_nodes/ComfyUI-YogurtNodes/yogurt_nodes/llm/ - the README documents the file format.

Common issues

Auth confusion is the theme. If you've set vertex true but left credentials empty, you get a cryptic Google error; if you've set vertex false but filled the Vertex fields, they're ignored. Check the toggle first, always. A blank text output usually means the request failed and retried - retry_count defaults to 1, so bump it if you're on flaky network. And since this sends your prompt to Google's servers, treat it like any cloud API: don't put anything you'd rather keep local into the prompt, and keep your API key out of shared workflows.