External Model - Language Model - Gemini

🛠

Frontier Text Reasoning via Google's Gemini Models

Read, reason over, or write text inside a CNAPS Studio flow with Google's Gemini models, including Gemini 4 Argon. Choose a supported model from the node's model dropdown. Processing runs on Google's servers through the Gemini API using your configured Google API key.

🔑 Requires a Google API key

Click to use Gemini on CNAPS Studio

What it does

This node connects your CNAPS Studio flows to Google's Gemini large language models through the Gemini API. Use it whenever a flow needs a model to read, reason over, or write text — interpreting the output of a vision model, drafting or enhancing prompts for image/video generators, summarizing, translating, or restructuring data. Because these run on Google's servers, you get frontier general-purpose reasoning without hosting a model locally.

Problem it solves

• Frontier reasoning without infrastructure — No local GPU or weights; the node calls Google's hosted models directly.
• Current Gemini models — Gemini 4 Argon is available in the text node; update existing flows to a supported model before running them again.
• Glue between vision and generation nodes — Turn raw model output into readable reports, or expand simple text into detailed generation prompts.

Which tier to pick

Select a supported model from the node's model dropdown and test it with your own inputs. Gemini 4 Argon supports up to 32K output tokens. When migrating from a retired model, review the system prompt and output token setting, then check that the response still meets your requirements for length, formatting, and content.

Configuring the Node

Basic Settings

Model name

System prompt or instruction — Tell the model how to behave. Examples:

• "Summarize the detection results in 2-3 sentences."
• "Convert these labels into a product description."
• "Translate the following text to French."

Advanced Options

Maximum tokens to generate (default: 2048)

• Controls the maximum response length. Increase this value for longer outputs, within the selected model's supported limit. Gemini 4 Argon supports up to 32K output tokens; this is the model's maximum, not the node's default setting.

Sampling temperature (default: 0.7, range: 0.0–1.0)

• Lower (0.0–0.3): Focused, consistent responses
• Higher (0.7–1.0): Creative, varied responses

Use Cases

Direction 1: Vision Model → Language Model (Analysis)

Best when vision output is complex or needs interpretation.

Example: Scene Segmentation → Report

Input - Image → Scene Segmentation (MaskFormer) → Language Model → Output - Text Viewer

System prompt:

Convert this segmentation data into a structured scene report. Include: 1) Setting type, 2) People count, 3) Key objects detected, 4) Overall scene summary. Use plain language, not raw data.

Direction 2: Language Model → Image Generation (Prompt Enhancement)

Input - Text Input → Language Model → Image Generation (Z-Image-Turbo) → Output - Image Viewer

System prompt:

You are an expert image prompt engineer. Transform the user's simple description into a detailed, vivid prompt optimized for AI image generation. Include: subject details, setting/environment, lighting, artistic style, mood/atmosphere, and composition. Output only the enhanced prompt, nothing else.

How Gemini compares

Gemini is one of three external LLM options in CNAPS Studio:

Compliance & Provenance

For more on how we classify models and mark outputs, see our AI Policy.

Run it now

Click to use Gemini