Gemini Text Generation

This is the node you'll actually live in. Gemini Text Generation is the pack's general-purpose LLM call: prompt in, STRING out, with a model dropdown spanning Gemini 3.5 Flash up through 3.1 Pro, plus the sampling knobs, JSON mode, and thinking-level control. Whether you're rewriting prompts, captioning, or extracting structured data, this is the default place to do it.

The model is picked right on the node (default gemini-3.5-flash - fast, cheap, plenty smart for prompt work), so you don't need a separate model selector. The client input is optional; key comes from Settings > ERPK > API Keys if you leave it unconnected.

The inputs that matter

• prompt - required. Multiline.
• temperature - 0.7 default, range 0–2. For deterministic extraction, drop it toward 0; for creative writing, push it up.
• max_tokens - 8192 default, up to 65536. Generous; cut it if you want faster/cheaper calls.
• top_p (0.95) and top_k (40) - nucleus and top-k sampling. Most people never touch them; 0.0 top_p disables, 0 top_k disables.
• stop_sequences - one per line; stop generation at these strings. Handy for trimming boilerplate.
• response_mime_type - default unless you set application/json to get JSON mode. That's the switch that turns this node into a structured-output node.
• response_schema - the JSON schema for structured output, used only with application/json. Empty = free-form JSON.
• thinking_level - none default; options up to high. This is the clever one: the node translates the same UI value into thinking_budget (an int) on Gemini 2.5 models or the thinking_level enum on 3.x, so the widget just works regardless of which model you pick.
• seed - reproducibility/cache control, -1 randomizes.

Output: a single response STRING - straight into Preview Anything or any downstream prompt input.

Install and gotchas

Part of the ERPK Collection:

cd ComfyUI/custom_nodes
git clone https://github.com/eRepublik-Labs/comfyui-nodes-erpk.git erpk
cd erpk && pip install -r requirements.txt

Restart (or ComfyUI Manager → search erpk), add a Google AI Studio API key in Settings > ERPK > API Keys. Paid per token; Gemini Flash models are cheap enough that this rarely stings, but Pro adds up.

Two things bite people. First, JSON mode is two fields, not one: you must set response_mime_type to application/json and give a response_schema (or leave it empty for free-form JSON) - setting only one does nothing useful. Second, thinking_level "none" is genuinely none; if you expected reasoning by default, flip it on. Otherwise you'll wonder why the "smart" model is giving you shallow answers.