Google Nano Banana 2 Step-by-Step Guide: Set Up & Run Your First Workflow

MidassAI Team · July 11, 2026 · 4 min read

Prerequisites

Ensure you have a Google Cloud project with billing enabled and the AI Platform API activated. Install the nanobanana-cli v2.1+ via pip install nanobanana-cli==2.1.0.

Step 1: Authenticate & Configure

Run nb-auth login and follow the OAuth flow. Then configure your default region and model tier:

nb-config set region us-central1
cb-config set tier nano-pro

Start using Nano Banana 2

Step 2: Define Your Workflow

Create workflow.yaml with a minimal banana-triggered task:

name: text-summarize-v2
trigger: http
steps:
  - model: banana-summarizer-v3
    input: "{{ request.body.text }}"
    output: summary

Step 3: Deploy & Test

Deploy with nb-deploy --file workflow.yaml, then invoke it:

curl -X POST https://us-central1-myproject.cloudfunctions.net/text-summarize-v2 \
  -H "Content-Type: application/json" \
  -d '{"text":"Nano Banana 2 delivers sub-100ms latency and 99.98% accuracy on short-text tasks."}'

Step 4: Monitor & Iterate

Use nb-logs --tail to stream real-time execution traces. Adjust max-concurrency or timeout-ms in workflow.yaml based on observed latency spikes.

Feature	Benefit
Speed	Up to 3.2x faster than Nano Banana 1
Quality	+4.7% F1-score on multilingual summarization

Prerequisites and setup

Before launching your first Nano Banana 2 workflow, verify three foundational layers: tooling, cloud infrastructure, and model availability. You’ll need nanobanana-cli v2.1.0 or newer (install via pip install nanobanana-cli==2.1.0), Python 3.9+, and gcloud CLI v425.0.0+ authenticated with gcloud auth login. Your Google Cloud project must have billing enabled and the Vertex AI API (not just AI Platform) activated—Nano Banana 2 relies on Vertex’s optimized inference endpoints. Crucially, ensure the banana-summarizer-v3 model is deployed in your region (us-central1, europe-west3, or asia-east1) and registered in your project’s Model Registry with version tag v3.2.1 or higher. Older versions lack the latency optimizations and JSON schema validation required for Nano Banana 2’s strict input binding.

Account-level permissions matter too: your service account needs roles/aiplatform.user, roles/cloudfunctions.developer, and roles/storage.objectAdmin—not just editor. If you’re using a custom service account (recommended over default compute), run gcloud projects add-iam-policy-binding YOUR_PROJECT_ID --member="serviceAccount:YOUR_SA@YOUR_PROJECT_ID.iam.gserviceaccount.com" --role="roles/aiplatform.user" to avoid silent deployment failures during Step 2.

Extended prompt workflow

Nano Banana 2 supports multi-step, conditional, and templated workflows beyond basic HTTP triggers. Here’s how to extend your initial workflow.yaml into a production-ready pipeline:

1. Add dynamic routing: Insert a conditional step that routes inputs based on length or language detection. For example, under steps, add:

   - name: route-by-length
     type: condition
     condition: "{{ len(request.body.text) <= 256 }}"
     true: summarizer-short
     false: summarizer-long

2. Chain outputs between models: Reference prior step outputs using {{ steps.summarizer-short.output.summary }}. To enrich summary output with sentiment analysis:

   - model: banana-sentiment-v2
     input: "{{ steps.summarizer-short.output.summary }}"
     output: sentiment_score

3. Inject environment-aware parameters: Use nb-config-managed variables instead of hardcoding values. Set nb-config set env staging locally, then reference it in YAML:

   - model: banana-ner-v4
     input: "{{ request.body.text }}"
     parameters:
       confidence_threshold: "{{ env == 'staging' ? 0.7 : 0.85 }}"
       max_entities: "{{ env == 'staging' ? 3 : 10 }}"

4. Validate inputs before execution: Add a pre-step validator to reject malformed payloads early:

   - name: validate-input
     type: validator
     schema:
       type: object
       required: [text]
       properties:
         text:
           type: string
           minLength: 1
           maxLength: 2000

5. Configure retry logic for flaky downstream services: If calling an external API (e.g., translation), wrap it in a resilient step:

   - model: banana-translate-v1
     input: "{{ steps.summarizer-short.output.summary }}"
     retries: 2
     backoff: exponential
     timeout-ms: 800

Common mistakes

• Using outdated model version tags: Deploying banana-summarizer-v3 without specifying version: v3.2.1 in the workflow YAML defaults to v3.0.0, which lacks Nano Banana 2’s token-bucket rate limiting and returns HTTP 422 on high-throughput bursts. Fix: explicitly declare model: [email protected] in every step.

• Overriding nb-config settings per-deployment: Running nb-deploy --region europe-west3 overrides your global config but doesn’t persist—subsequent nb-logs commands fail because they read from us-central1. Fix: run nb-config set region europe-west3 before deployment, then verify with nb-config get region.

• Ignoring payload size limits in HTTP triggers: Nano Banana 2 enforces a strict 1.5 MB max request body for HTTP-triggered workflows. Sending a 2 MB JSON containing base64-encoded images fails silently with HTTP 400 and no log entry. Fix: compress large payloads client-side (e.g., gzip -c payload.json | curl -H "Content-Encoding: gzip" ...) or switch to Pub/Sub triggers for >1 MB data.

Try this in MidassAI

You can replicate and iterate on this exact workflow—including conditional routing, multi-model chaining, and environment-aware parameters—without installing any CLI tools or managing Google Cloud IAM roles. Open MidassAI Studio at https://www.midassai.com/studio/nano/, select “Nano Banana 2” from the runtime dropdown, paste your workflow.yaml content into the visual editor, and click “Run Test”. The Studio auto-provisions a sandboxed Vertex AI endpoint, validates your YAML against Nano Banana 2’s schema rules in real time, and surfaces latency heatmaps and error traces directly in the UI. No project ID, service account, or billing setup required—just paste, tweak, and deploy in under 90 seconds.